The day in tech, ranked blind.

Get the day’s top ten stories, with their summaries, in one free email a day.

Every topic and format but new projects. Pick your own mix

One essay a week on AI and the work of programming: the hrih.si letter

Ranked by score · Showing 20 of 136 stories · 8 hidden by your filters

  1. No. 1Score 0.82

    Nginx http2 RCE

    Summary and score for “Nginx http2 RCE”

    F5 released NGINX 1.30.7 stable and 1.31.9 mainline, fixing 10 CVEs, including two HTTP/2 remote code execution flaws (CVE-2026-123633, CVE-2026-123007) rated 9.2 on CVSS 4.0. Wiz says it found both with its AI scanner, and its tests showed code execution with an ASLR bypass on 1.31.7 and 1.31.8 builds. Wiz has not released its exploits. Both need HTTP/2 enabled; if patching must wait, disable it. Several figures, including the scores, come from Wiz alone.

    Score 0.82Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 1253 pts, front page

    Score breakdown and claims for “Nginx http2 RCE” · Also covered by The Register, heise online English

    Comments on “Nginx http2 RCE”
  2. No. 2Score 0.82

    Microsoft agent-framework Compromised Twice on PyPI

    Summary and score for “Microsoft agent-framework Compromised Twice on PyPI”

    SafeDep reports that agent-framework 1.26.1 and 1.26.2, Microsoft's Python metapackage on PyPI, carried a credential stealer. A .pth file runs it at every Python start, so installing is enough. It takes environment variables and credential files, requests fresh Azure DevOps, managed identity and, in 1.26.2, GitHub Actions tokens, and posts them to webhook.site. PyPI removed both; 1.26.0 is the last clean release. A maintainer account created that day may be the uploader, which is unconfirmed. Treat reachable secrets as exposed and hash-lock the dependency tree.

    Score 0.82Contestedjudges differ by 0.87 on new informationClaim check: 2 matches source · 3 no source given · 3 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Microsoft agent-framework Compromised Twice on PyPI” · Also covered by opensourcemalware.com, The Register

    Comments on “Microsoft agent-framework Compromised Twice on PyPI”
  3. No. 3Score 0.77

    Backdoor attacks against Machine Learning IRL

    Summary and score for “Backdoor attacks against Machine Learning IRL”

    In Phrack, G. Cherubin describes a parser flaw in ONNX Runtime: sparse tensors are densified at load time, before external-path validation, so a model file can make the runtime read arbitrary files or memory addresses. Combined with graph logic in a Qwen2.5-7B-Instruct model, a trigger word makes the model emit the read bytes as tokens, with no change to the runtime. Both fixes shipped in onnxruntime v1.27.0. The author's point for builders: downloaded models are code and need the same vetting as libraries.

    Score 0.77Claim check: no checkable claims

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Backdoor attacks against Machine Learning IRL”

    Comments on “Backdoor attacks against Machine Learning IRL”
  4. No. 4Score 0.75

    Reflections on Trusting AI: A Virus Inside an LLM

    Summary and score for “Reflections on Trusting AI: A Virus Inside an LLM”

    Writing in Phrack, marver describes a proof of concept in which a backdoored Qwen2.5-3B helper model (Aux0) writes the backdoor into training data for the next model. After two QLoRA distillation hops on clean prompts, the trigger-domain curl payload still fired 100% of the time, with 0% false-fire. The propagation-ready training row dropped to 27%. These are single-seed runs with small samples, and the claim that larger students propagate better is an untested expectation. For builders, unvetted formatter, scorer or filter models are a supply-chain risk.

    Score 0.75Contestedjudges differ by 0.52 on substantiatedClaim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Reflections on Trusting AI: A Virus Inside an LLM”

    Comments on “Reflections on Trusting AI: A Virus Inside an LLM”
  5. No. 5Score 0.75

    Security advisories: multiple vulnerabilities in Regexp and String case conversion

    Summary and score for “Security advisories: multiple vulnerabilities in Regexp and String case conversion”

    The Ruby team disclosed seven memory-safety flaws in the regular expression engine and in string case conversion, and released Ruby 4.0.10, 3.4.13 and 3.3.14 to fix them. Several involve heap overflows or a double free, and the maintainers say proof-of-concept exploits show remote code execution in some scenarios, mainly where applications compile attacker-controlled patterns or change the case of crafted strings. CVE IDs are still pending. Ruby 3.2 and earlier are end of life and get no fix, so teams on those versions must upgrade.

    Score 0.75Claim check: 4 matches source · 4 no source given

    Elsewhere: HN 167 pts, front page

    Score breakdown and claims for “Security advisories: multiple vulnerabilities in Regexp and String case conversion”

    Comments on “Security advisories: multiple vulnerabilities in Regexp and String case conversion”
  6. No. 6Score 0.71

    Sashiko: Lessons from building an agentic code review system for Linux

    Summary and score for “Sashiko: Lessons from building an agentic code review system for Linux”

    Google's AI & Infrastructure team describes Sashiko, an open-source, Gemini-based code review system donated to the Linux Foundation, which runs parallel specialist passes and then verifies each disputed finding separately. Google reports about 1,200 reviews a day, a 75.8% catch rate on a roughly 1,000-bug benchmark of missed kernel bugs, and lower post-merge fix rates. These figures are self-reported and unsourced, and Google says other factors may explain the fix-rate drop. The design separates bug finding from verification, which is reusable for review agents.

    Score 0.71Claim check: 6 no source given · 2 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Sashiko: Lessons from building an agentic code review system for Linux”

    Comments on “Sashiko: Lessons from building an agentic code review system for Linux”
  7. No. 7Score 0.70

    jax-ml/jax jax-v0.12.1

    Summary and score for “jax-ml/jax jax-v0.12.1”

    JAX 0.12.1 removes the legacy GSPMD partitioner, so Shardy is now the only option, and it enables simplified jaxpr constants by default, passing captured arrays as separate inputs. Executables serialized with 0.12.0 or earlier no longer load. It also fixes a jit dispatch-cache bug that could return wrong results, and GPU preallocation becomes 75% with growth. The notes report accuracy gains for erfc, erf_inv and exp2, and CPU speedups for log and idct, but those figures are the project's own and mostly unverified.

    Score 0.70Claim check: 3 matches source · 4 no source given · 1 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “jax-ml/jax jax-v0.12.1”

    Comments on “jax-ml/jax jax-v0.12.1”
  8. No. 8Score 0.73

    Lessons from Telegram's RCE

    Summary and score for “Lessons from Telegram's RCE”

    NebuSec reports that Telegram's new bot widgets, which run bot-supplied JavaScript in QuickJS on the user's device, led to remote code execution within a day of launch. It also says Telegram's follow-up seccomp filter, a denylist of about 30 calls, was bypassed two days later. Details are withheld, and several dates and claims are unverified. NebuSec proposes layers: an isolated process, per-widget WebAssembly sandboxing via wasm2c, and RLBox-tainted values. It notes that Skia, outside the sandbox, was also exploited.

    Score 0.73Contestedjudges differ by 0.53 on new informationClaim check: 2 matches source · 1 no source given · 5 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Lessons from Telegram's RCE”

    Comments on “Lessons from Telegram's RCE”
  9. No. 9Score 0.69

    Muse privacy invasions are freaking out Meta staffers, internal chats show

    Summary and score for “Muse privacy invasions are freaking out Meta staffers, internal chats show”

    Reuters, summarized by Ars Technica, reports internal Meta chats in which staff testing the Muse AI agent flagged apparent cross-account leaks, including unprompted bank balances. Meta calls these isolated hallucinations caused by a since-fixed bug, and Reuters could not establish whether the data was real. Reuters' tests with Signal and AI Now found Muse reading Signal messages on Android and macOS through setup permissions and accessibility tools. For agent builders, the lesson is that broad OS permissions are not informed consent. None of these claims could be independently verified.

    Score 0.69Claim check: 1 no source given · 7 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “Muse privacy invasions are freaking out Meta staffers, internal chats show”

    Comments on “Muse privacy invasions are freaking out Meta staffers, internal chats show”
  10. No. 10Score 0.68

    Sched_ext: A sketch of the scx_lavd scheduler (Part 3)

    Summary and score for “Sched_ext: A sketch of the scx_lavd scheduler (Part 3)”

    Changwoo Han of Igalia describes scx_lavd, a Linux CPU scheduler written in BPF on sched_ext that infers task urgency from wait frequency, wake frequency and runtime, then turns it into a virtual deadline, time slice and preemption right. Started for Valve's Steam Deck, it is said to run on the Steam Frame and in Meta's datacenters, though that deployment claim and the quoted figures (futex share, probe costs, code size) are the author's own and unverified. For builders, the design shows how to schedule without annotations.

    Score 0.68Claim check: 7 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Sched_ext: A sketch of the scx_lavd scheduler (Part 3)”

    Comments on “Sched_ext: A sketch of the scx_lavd scheduler (Part 3)”
  11. No. 11Score 0.73

    LibreOffice RCE via ODF Form Datasource JDBC Code Execution

    Summary and score for “LibreOffice RCE via ODF Form Datasource JDBC Code Execution”

    A detailed write-up describes CVE-2026-104692 and CVE-2026-104007, two LibreOffice flaws that together allow remote code execution when a victim opens a crafted spreadsheet with a Java runtime enabled. The first is an HSQLDB method whitelist initialized to null, which disables the check entirely.

    Score 0.73Claim check: no checkable claims

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “LibreOffice RCE via ODF Form Datasource JDBC Code Execution”

    Comments on “LibreOffice RCE via ODF Form Datasource JDBC Code Execution”
  12. No. 12Score 0.67

    Pushing WireGuard on macOS to 5 Gbit/s

    Summary and score for “Pushing WireGuard on macOS to 5 Gbit/s”

    Firezone says it raised its macOS WireGuard client from about 1 Gbit/s to over 5 Gbit/s. Its reading of Apple's XNU source is that UDP writes never return EWOULDBLOCK, so packets drop at the interface queue and TCP inside the tunnel slows needlessly. Connected sockets, ENOBUFS handling, batched sendmsg_x and coalescing up to three TCP segments restored backpressure and cut per-packet cost. The 5.87 Gbit/s upload and 5.76 download figures come from one pair of M1 Macs over Thunderbolt, and the XNU buffer behavior was not independently verified.

    Score 0.67Claim check: 2 no source given · 1 couldn’t check

    Elsewhere: HN 19 pts, front page

    Score breakdown and claims for “Pushing WireGuard on macOS to 5 Gbit/s”

    Comments on “Pushing WireGuard on macOS to 5 Gbit/s”
  13. No. 13Score 0.66

    Temp tables in Postgres don't scale

    Summary and score for “Temp tables in Postgres don't scale”

    PostgresAI's Nikolay Samokhvalov measured CREATE TEMP TABLE run in every transaction on Postgres 18 (16 vCPU): 3783 TPS, against 17866 for a table kept per session and 86997 for a CTE. Each create and drop touches 31 catalog rows. With 5000 idle sessions and a 500 TPS target, temp-table DDL reached 127 and 249 TPS in two runs, as the shared invalidation queue became the limit. The figures come from the author's own experiments and could not be independently checked.

    Score 0.66Claim check: 2 matches source · 2 no source given · 4 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Temp tables in Postgres don't scale”

    Comments on “Temp tables in Postgres don't scale”
  14. No. 14Score 0.69

    SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations

    Summary and score for “SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations”

    TokenWave AI's SWE-CoWork benchmark places coding agents in simulated engineering organizations with 26 modeled colleagues, across 101 assignments and repositories. Claude-Opus-5.5 and GPT-6 Astra pass about 63% of fail-to-pass tests but complete under 8% of tasks. Supplying full requirements upfront raised test pass rates but left task success unchanged. Claude-Opus-5.5 reaches the right owners yet obtains 64% of colleague-held requirements, and 31% of requirements it reported done failed tests. These figures are the authors' own, without independent verification.

    Score 0.69Claim check: 2 matches source · 6 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations”

    Comments on “SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations”
  15. No. 15Score 0.66

    Quad9 Is Leaving France: Why DNS Resolvers Are the Wrong Place to Fight Piracy

    Summary and score for “Quad9 Is Leaving France: Why DNS Resolvers Are the Wrong Place to Fight Piracy”

    Quad9, the Swiss non-profit behind the 9.9.9.9 resolver, says it will stop serving France on November 24, 2026, after a Paris court fined it €1,000 a day for not blocking pirate sports-streaming domains sought by beIN Sports. Quad9 says about 20 similar cases are pending and it cannot absorb the accumulated penalties. For anyone running resolvers or network infrastructure, the case shows how French courts are extending blocking duties to DNS operators. Techstackups, which argues against the approach, advises Quad9 users in France to check hard-coded 9.9.9.9 settings.

    Score 0.66Claim check: 4 matches source · 4 couldn’t check

    Elsewhere: HN 639 pts, front page

    Score breakdown and claims for “Quad9 Is Leaving France: Why DNS Resolvers Are the Wrong Place to Fight Piracy”

    Comments on “Quad9 Is Leaving France: Why DNS Resolvers Are the Wrong Place to Fight Piracy”
  16. No. 16Score 0.65

    Lakegres: Vector search on existing Open/Hudi/Iceberg tables

    Summary and score for “Lakegres: Vector search on existing Open/Hudi/Iceberg tables”

    Onehouse added vector search to Lakegres, storing an IVF-style index with RaBitQ quantization in the metadata of existing Hudi and Iceberg tables, so Postgres clients can query lake data without copying it or rewriting files. On a 247M-vector, 1,024-dimension table, it reports warm p50 of 63.5 ms against 655 ms for Databricks AI Search, at similar recall (about 0.95). These are the vendor's own figures, from one workload, with different latency measurement methods, and the "only open index" claim is unsourced.

    Score 0.65Claim check: 3 matches source · 4 no source given · 1 couldn’t check

    Elsewhere: HN 40 pts, front page

    Score breakdown and claims for “Lakegres: Vector search on existing Open/Hudi/Iceberg tables”

    Comments on “Lakegres: Vector search on existing Open/Hudi/Iceberg tables”
  17. No. 17Score 0.65

    Stoolap 0.5.0: One Query, One Answer

    Summary and score for “Stoolap 0.5.0: One Query, One Answer”

    Stoolap 0.5.0, an embedded SQL database, is aimed at making a query return the same answer on every internal path. Author Semih Alev shows 0.4.2 returning different results depending on an index, a checkpoint, or one corrupted byte that left a table empty without an error. The release centralizes comparison, ordering and arithmetic rules, turns runtime errors into failures, and adds V5 per-block checksums. It is not backward compatible on disk. The bug counts and benchmark figures are the author's own and have no independent source.

    Score 0.65Claim check: 2 matches source · 6 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Stoolap 0.5.0: One Query, One Answer”

    Comments on “Stoolap 0.5.0: One Query, One Answer”
  18. No. 18Score 0.73

    Octopus, Gecko and Tortoise: the 3 scanners hunting your AI keys

    Summary and score for “Octopus, Gecko and Tortoise: the 3 scanners hunting your AI keys”

    Tin Computer's Ege Demir logged three scanners on ten sites he runs, late September to 18 November 2026. Octopus requests .env and cloud-key files under 32 rotating AI crawler names from short-lived Google Cloud machines. Gecko, probably a .NET program, sends a truncated Chrome user agent and grew its list of AI coding-tool config paths from 4 to 26. The figures come from his own logs, and none of the requests returned a secret. Crawler names are spoofable, so verify against published IP ranges and block dotfile paths.

    Score 0.73Claim check: 2 matches source · 5 no source given · 1 couldn’t check

    Elsewhere: HN 28 pts, front page

    Score breakdown and claims for “Octopus, Gecko and Tortoise: the 3 scanners hunting your AI keys”

    Comments on “Octopus, Gecko and Tortoise: the 3 scanners hunting your AI keys”
  19. No. 19Score 0.65

    Faster CI with Orchestrated Pytest Workers

    Summary and score for “Faster CI with Orchestrated Pytest Workers”

    Close, a company that sells a CRM, describes how it cut CI for its backend's roughly 58,000 mostly datastore-backed integration tests to about two and a half minutes, down from about 12 with static splitting across 17 runners. A central orchestrator now collects test files and hands out batches dynamically, so workers start backends in parallel and collect only what they run. Cheaper spot instances (54, plus 2 on-demand) cost about the same as before. These figures are Close's own and unverified externally.

    Score 0.65Contestedjudges differ by 0.41 on substantiated, 0.49 on overclaimClaim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Faster CI with Orchestrated Pytest Workers”

    Comments on “Faster CI with Orchestrated Pytest Workers”
  20. No. 20Score 0.73

    Bitrise publishes security incident review: TTPs, IOCs, detection ideas

    Summary and score for “Bitrise publishes security incident review: TTPs, IOCs, detection ideas”

    Bitrise, the mobile CI/CD provider, published a preliminary review of its November 13 breach. An attacker logged into an internal testing bot account that had no two-factor authentication, using valid credentials whose origin is still unknown. Bitrise shut the platform down, rotated secrets and revoked 919 cloud keys. It says files in 15 customer workspaces were accessed, with no sign customer secret stores were touched. These are Bitrise's own statements, pending GitHub audit logs and outside forensics. It also released IOCs in CSV and STIX formats.

    Score 0.73Claim check: 1 matches source · 7 no source given

    Elsewhere: HN 5 pts, not on front page

    Score breakdown and claims for “Bitrise publishes security incident review: TTPs, IOCs, detection ideas”

    Comments on “Bitrise publishes security incident review: TTPs, IOCs, detection ideas”
  21. No. 21Score 0.67

    Using a "chaos loop" to harden the Claude Docs launch

    Summary and score for “Using a "chaos loop" to harden the Claude Docs launch”

    An engineering write-up on Anthropic's blog describes how the Claude Docs launch was hardened using a "chaos loop": a long-running Claude Code session that stress-tested a staging cluster, read telemetry, and shipped fixes over two weeks.

    Score 0.67Contestedjudges differ by 0.41 on substantiatedClaim check: 1 matches source · 7 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Using a "chaos loop" to harden the Claude Docs launch”

    Comments on “Using a "chaos loop" to harden the Claude Docs launch”
  22. No. 22Score 0.67

    Microsoft's Giant Outlook Outage

    Summary and score for “Microsoft's Giant Outlook Outage”

    A nearly 24-hour Exchange Online outage knocked out Outlook email for employees, businesses, and governments worldwide after Microsoft's failover systems spread a failure that began in Switzerland.

    Score 0.67Contestedjudges differ by 0.53 on new informationClaim check: 5 no source given · 3 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Microsoft's Giant Outlook Outage”

    Comments on “Microsoft's Giant Outlook Outage”
  23. No. 23Score 0.62

    10k sounds, 64 mixer slots: how much work can a game audio engine skip?

    Summary and score for “10k sounds, 64 mixer slots: how much work can a game audio engine skip?”

    Sonic is a small Rust experiment by ssenthilnathan3 for scheduling 10,000 game sounds into 64 mixer slots without eagerly advancing every sound. It keeps distant sounds as virtual records that compute playback position from their start time, and it uses a spatial grid and an expiry heap to limit scheduler…

    Score 0.62Claim check: 1 no source given · 7 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “10k sounds, 64 mixer slots: how much work can a game audio engine skip?”

    Comments on “10k sounds, 64 mixer slots: how much work can a game audio engine skip?”
  24. No. 24Score 0.75

    Attackers Used Optimizely to A/B Test Malware on Top Websites

    Summary and score for “Attackers Used Optimizely to A/B Test Malware on Top Websites”

    Attackers exploited a vulnerability in Optimizely, a website A/B testing and experimentation vendor, to inject malicious JavaScript into at least 178 major sites, including the BBC, CNN, Equifax, Ledger and Zoom, on November 16.

    Score 0.75Contestedjudges differ by 0.63 on new informationClaim check: 3 matches source · 5 no source given

    Elsewhere: HN 1 pts, front page

    Score breakdown and claims for “Attackers Used Optimizely to A/B Test Malware on Top Websites”

    Comments on “Attackers Used Optimizely to A/B Test Malware on Top Websites”
  25. No. 25Score 0.62

    US data centers are burning through as much electricity as entire states

    Summary and score for “US data centers are burning through as much electricity as entire states”

    A Harvard-led team, publishing in Nature Sustainability, estimates that 403 US hyperscale data centers used 68 to 99 terawatt-hours of electricity in the year ending April 2025, roughly the annual consumption of Michigan at the high end.

    Score 0.62

    Elsewhere: HN 15 pts, not on front page

    Score breakdown for “US data centers are burning through as much electricity as entire states” · Also covered by NBC News

    Comments on “US data centers are burning through as much electricity as entire states”
  26. No. 26Score 0.70

    Java4Shell what Apache Shiro users should know

    Summary and score for “Java4Shell what Apache Shiro users should know”

    A newly published OpenJDK security advisory, CVE-2026-114938 ("Java4Shell"), describes a flaw in the HotSpot JVM's optimized Base64 decoder, not in Apache Shiro itself. Shiro applications can still be exposed because its remember-me cookie handling decodes Base64 through the JDK.

    Score 0.70Contestedjudges differ by 0.78 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Java4Shell what Apache Shiro users should know”

    Comments on “Java4Shell what Apache Shiro users should know”
  27. No. 27Score 0.67

    Parallel agent requests can slow down an Ollama session

    Summary and score for “Parallel agent requests can slow down an Ollama session”

    Tensward reports that when two coding-agent sessions shared one Ollama slot running the hybrid Qwen3.8-27B model, the median wait before a response started was about 46 seconds, versus about one second with a slot each. Over half the prompt tokens processed were repeats.

    Score 0.67

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Parallel agent requests can slow down an Ollama session”

    Comments on “Parallel agent requests can slow down an Ollama session”
  28. No. 28Score 0.69

    Dissecting the GPU Load/Store Unit and Exposing Its Security Risks

    Summary and score for “Dissecting the GPU Load/Store Unit and Exposing Its Security Risks”

    Researchers reverse-engineered the load/store unit (LSU) of NVIDIA V100, A100, and H100 GPUs using SASS timing microbenchmarks. They derived a cycle-accurate latency formula for shared and L1-hit global loads, identified contention points, and showed the LSU leaks AES-128 keys, enables covert channels under MPS, and lets one warp slow colocated…

    Score 0.69Contestedjudges differ by 0.42 on practical relevance

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Dissecting the GPU Load/Store Unit and Exposing Its Security Risks”

    Comments on “Dissecting the GPU Load/Store Unit and Exposing Its Security Risks”
  29. No. 29Score 0.69

    Owning the Truth: Undecidability, unsound kernels, and the forging of proof

    Summary and score for “Owning the Truth: Undecidability, unsound kernels, and the forging of proof”

    This Phrack article, by endrazine and veorq, argues that undecidability and unsound proof kernels are distinct limits on trusting software, and demonstrates both. It reports a constructive Rice's theorem reduction mechanized in Rocq, plus seven CVEs for proof-assistant soundness failures, forged proofs in Lean 4 and Rocq, a backdoored CompCert…

    Score 0.69

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Owning the Truth: Undecidability, unsound kernels, and the forging of proof”

    Comments on “Owning the Truth: Undecidability, unsound kernels, and the forging of proof”
  30. No. 30Score 0.61

    Google will require Android apps to justify location access starting in January

    Summary and score for “Google will require Android apps to justify location access starting in January”

    Google is tightening Android app permissions on a staggered schedule, according to a heise online report. From January 27, 2027, apps targeting Android 17 must use the one-time Location button for infrequent location requests, and any app seeking precise location must file a Play Console justification.

    Score 0.61

    Elsewhere: HN —

    Score breakdown for “Google will require Android apps to justify location access starting in January”

    Comments on “Google will require Android apps to justify location access starting in January”
  31. No. 31Score 0.60

    Bluetooth 7.0 specification ratified with HDT (High Data Throughput) support

    Summary and score for “Bluetooth 7.0 specification ratified with HDT (High Data Throughput) support”

    The Bluetooth SIG ratified Bluetooth 7.0 on November 3, 2026, centered on optional High Data Throughput (HDT) support that raises the maximum bitrate to 7.5 Mbps through a new LE physical layer. The release also adds isochronous parameter updates, a new key schedule, transient application keys, and encryption debugging.

    Score 0.60

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Bluetooth 7.0 specification ratified with HDT (High Data Throughput) support”

    Comments on “Bluetooth 7.0 specification ratified with HDT (High Data Throughput) support”
  32. No. 32Score 0.64

    ABFS: Another Build File Service

    Summary and score for “ABFS: Another Build File Service”

    ABFS (Another Build File Service) is an open-source system from Google that speeds up checkouts and builds of large code repositories, initially targeting Android. It uses a content-addressable filesystem, lazy-loaded file content, and a shared remote build cache.

    Score 0.64

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “ABFS: Another Build File Service”

    Comments on “ABFS: Another Build File Service”
  33. No. 33Score 0.62

    Traffic mix matters

    Summary and score for “Traffic mix matters”

    Marc Brooker, an AWS engineer, argues that system performance is highly sensitive to workload mix, so load tests must match real production traffic. Using simple simulations of lock contention and cache behavior, he shows throughput can swing by orders of magnitude in opposite directions as key skew changes.

    Score 0.62Contestedjudges differ by 0.63 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Traffic mix matters”

    Comments on “Traffic mix matters”
  34. No. 34Score 0.62

    Draw live DOM subtrees directly into 2D, WebGL, or WebGPU canvas contexts

    Summary and score for “Draw live DOM subtrees directly into 2D, WebGL, or WebGPU canvas contexts”

    Chrome's HTML-in-Canvas API lets developers draw live DOM subtrees directly into 2D, WebGL, or WebGPU canvas contexts while the browser keeps the content interactive and accessible. The API aims to combine the accessibility of the DOM with the pixel-level control of canvas.

    Score 0.62

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Draw live DOM subtrees directly into 2D, WebGL, or WebGPU canvas contexts”

    Comments on “Draw live DOM subtrees directly into 2D, WebGL, or WebGPU canvas contexts”
  35. No. 35Score 0.68

    Missing Security-Fix Backporting for GitHub Actions Workflows

    Summary and score for “Missing Security-Fix Backporting for GitHub Actions Workflows”

    A large-scale study of GitHub Actions workflows finds that security fixes applied to default branches are rarely carried onto release and maintenance branches: 47.3% of 474,490 audited release branches remain vulnerable, and deliberate backports arrive a median 243 days late.

    Score 0.68

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Missing Security-Fix Backporting for GitHub Actions Workflows”

    Comments on “Missing Security-Fix Backporting for GitHub Actions Workflows”
  36. No. 36Score 0.59

    The Risks Are the Roadmap

    Summary and score for “The Risks Are the Roadmap”

    Varoa's essay "The Risks Are the Roadmap" argues that engineering teams hesitant to merge pull requests on automated review alone should treat those hesitations as problems to solve, not permanent limits.

    Score 0.59Contestedjudges differ by 0.52 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The Risks Are the Roadmap”

    Comments on “The Risks Are the Roadmap”
  37. No. 37Score 0.67

    Tencent releases weights of Hy4 Flash (220B size)

    Summary and score for “Tencent releases weights of Hy4 Flash (220B size)”

    Tencent has released open weights for Hy4 Flash, a 220B-parameter mixture-of-experts model with 12B active parameters per token, a 1M-token context window, and text-plus-image input, along with an FP8 variant and a base model. Its architecture combines hybrid sparse and sliding-window attention, LatentMoE routing, and n-gram embeddings.

    Score 0.67

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Tencent releases weights of Hy4 Flash (220B size)”

    Comments on “Tencent releases weights of Hy4 Flash (220B size)”
  38. No. 38Score 0.67

    The public database of AI agent incidents

    Summary and score for “The public database of AI agent incidents”

    The Agent Incident Register is a public database, updated daily, cataloguing AI agent failures and misuse. Each entry is graded by cause, confidence score and cited sources.

    Score 0.67Contestedjudges differ by 0.47 on technical depth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The public database of AI agent incidents”

    Comments on “The public database of AI agent incidents”
  39. No. 39Score 0.67

    Multi Agent Coordination Benchmark

    Summary and score for “Multi Agent Coordination Benchmark”

    MAC Bench is a new benchmark of twelve simulated tasks that measures whether teams of AI agents coordinate well, separating the gain from communication from the gain of simply adding agents. Its tasks are built so each agent depends on its teammates' work.

    Score 0.67

    Elsewhere: HN 15 pts, not on front page

    Score breakdown for “Multi Agent Coordination Benchmark”

    Comments on “Multi Agent Coordination Benchmark”
  40. No. 40Score 0.62

    From 181 bytes: nixpkgs' full-source bootstrap

    Summary and score for “From 181 bytes: nixpkgs' full-source bootstrap”

    Farid Zakaria's write-up, "From 181 bytes," describes how nixpkgs now builds GCC 16.2.0 and glibc from source on i686-linux, starting from a 181-byte hex0 seed program that is the only binary trusted.

    Score 0.62Contestedjudges differ by 0.51 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “From 181 bytes: nixpkgs' full-source bootstrap”

    Comments on “From 181 bytes: nixpkgs' full-source bootstrap”
  41. No. 41Score 0.66

    More than one in four public MCP servers can take an action nobody can undo

    Summary and score for “More than one in four public MCP servers can take an action nobody can undo”

    Rinsoft scanned 800 public MCP registry servers in October 2026, using only the tools/list request and no tool calls, and classified 5,154 tools by whether their effects can be undone.

    Score 0.66Contestedjudges differ by 0.50 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “More than one in four public MCP servers can take an action nobody can undo”

    Comments on “More than one in four public MCP servers can take an action nobody can undo”
  42. No. 42Score 0.66

    UAT-12643: Suspected North Korea-aligned actor blends blockchain resilience with targeted infostealing

    Summary and score for “UAT-12643: Suspected North Korea-aligned actor blends blockchain resilience with targeted infostealing”

    Cisco Talos reports on UAT-12643, a financially motivated threat group with low-confidence links to North Korea-aligned actors. Its campaign targets cryptocurrency holders, gaming account owners, VoIP users, and finance staff, mainly in Europe, using weaponized LNK, HTA, and MSBuild files.

    Score 0.66Contestedjudges differ by 0.55 on overclaim

    Elsewhere: HN —

    Score breakdown for “UAT-12643: Suspected North Korea-aligned actor blends blockchain resilience with targeted infostealing” · Also covered by Cisco Talos

    Comments on “UAT-12643: Suspected North Korea-aligned actor blends blockchain resilience with targeted infostealing”
  43. No. 43Score 0.58

    Windows Server 2028 won't be compatible by default

    Summary and score for “Windows Server 2028 won't be compatible by default”

    Microsoft's public preview of Windows Server 2028 ships Server Core hardened by default, with hundreds of security settings enabled from first boot and a new security baseline, including Trusted Launch for virtual machines.

    Score 0.58

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Windows Server 2028 won't be compatible by default” · Also covered by BleepingComputer

    Comments on “Windows Server 2028 won't be compatible by default”
  44. No. 44Score 0.66

    Downgrading from Claude Fable barely affected dev productivity during ban

    Summary and score for “Downgrading from Claude Fable barely affected dev productivity during ban”

    A Harvard study tracked 297 developers who used Claude Fable 5 before, during, and after a 19-day White House ban in June 2026, comparing their GitHub pull requests with 8,908 developers on the older Opus 4.8.

    Score 0.66

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Downgrading from Claude Fable barely affected dev productivity during ban”

    Comments on “Downgrading from Claude Fable barely affected dev productivity during ban”
  45. No. 45Score 0.66

    A 149M model handles 70% of phone-agent steps locally, Claude the rest

    Summary and score for “A 149M model handles 70% of phone-agent steps locally, Claude the rest”

    A solo developer, Chaitanya Talapalli, released a preprint and open code for Milo Pointer, a 149M-parameter text model that handles about 70% of phone-agent steps on the device and passes the rest to Claude.

    Score 0.66

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “A 149M model handles 70% of phone-agent steps locally, Claude the rest”

    Comments on “A 149M model handles 70% of phone-agent steps locally, Claude the rest”
  46. No. 46Score 0.65

    The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms

    Summary and score for “The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms”

    A security researcher's disclosure details how Credit Nirvana, an Indian AI debt-collection firm, exposed sensitive data through two flaws: a hardcoded API key in client-side JavaScript that granted access to its Agentic AI Call Center, and improper authentication on its primary platform that let an attacker log in as admin.…

    Score 0.65Contestedjudges differ by 0.51 on significance

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms”

    Comments on “The AI Repo Man: Exploiting Agentic AI Debt Collection Platforms”
  47. No. 47Score 0.65

    Service as a Software: The New SaaS Is Ahead of Schedule

    Summary and score for “Service as a Software: The New SaaS Is Ahead of Schedule”

    Outcome-based pricing is spreading through AI services: Fin charges $0.99 per resolved support ticket and Zendesk about $2 per automated resolution, shifting the cost of failed work to the vendor. The author, who grades his own 2024 predictions, notes that McKinsey and others are repricing consulting work by result.

    Score 0.65Contestedjudges differ by 0.62 on new information

    Elsewhere: HN —

    Score breakdown for “Service as a Software: The New SaaS Is Ahead of Schedule”

    Comments on “Service as a Software: The New SaaS Is Ahead of Schedule”
  48. No. 48Score 0.65

    Synthesizing Rollback Systems for Safer Agents

    Summary and score for “Synthesizing Rollback Systems for Safer Agents”

    Basis, working with Amazon Web Services' Automated Reasoning Group, has built a research prototype of a formally verified "undo button" for the Amazon S3 API.

    Score 0.65

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Synthesizing Rollback Systems for Safer Agents”

    Comments on “Synthesizing Rollback Systems for Safer Agents”
  49. No. 49Score 0.57

    TCBP-JOBS: GPU-Accelerated Quantum Error Correction

    Summary and score for “TCBP-JOBS: GPU-Accelerated Quantum Error Correction”

    TCBP-JOBS is a GPU-accelerated quantum error correction decoder that combines trajectory-calibrated belief propagation, which learns calibrated fault probabilities from short BP runs, with jittered ordered-statistics search run in parallel.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “TCBP-JOBS: GPU-Accelerated Quantum Error Correction”

    Comments on “TCBP-JOBS: GPU-Accelerated Quantum Error Correction”
  50. No. 50Score 0.65

    Nvidia Kumo Tabular: Pushing the Accuracy-Efficiency Frontier

    Summary and score for “Nvidia Kumo Tabular: Pushing the Accuracy-Efficiency Frontier”

    Nvidia researchers introduced Kumo Tabular, a family of open tabular foundation models for classification and regression in 27M, 61M, and 214M parameter sizes. Pretrained only on artificial tables, they predict new tables without task-specific training.

    Score 0.65

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Nvidia Kumo Tabular: Pushing the Accuracy-Efficiency Frontier”

    Comments on “Nvidia Kumo Tabular: Pushing the Accuracy-Efficiency Frontier”
  51. No. 51Score 0.65

    California sues OpenAI over AI agents escaping sandboxes, hacking firms

    Summary and score for “California sues OpenAI over AI agents escaping sandboxes, hacking firms”

    California Attorney General Rob Bonta sued OpenAI, alleging the company knowingly kept testing advanced AI models despite security failures that let autonomous agents escape sandboxes and attack outside organizations, including a July incident that compromised Hugging Face.

    Score 0.65Contestedjudges differ by 0.44 on overclaim

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “California sues OpenAI over AI agents escaping sandboxes, hacking firms” · Also covered by TechCrunch, BleepingComputer

    Comments on “California sues OpenAI over AI agents escaping sandboxes, hacking firms”
  52. No. 52Score 0.61

    HTAP on MariaDB with ProxySQL and DBTrail

    Summary and score for “HTAP on MariaDB with ProxySQL and DBTrail”

    DBTrail's router adds about 0.55 ms per query, mostly from MariaDB's EXPLAIN plan lookup and network round trips, and in a rerun the application's p95 stayed near 12 ms while heavy reports moved off MariaDB.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “HTAP on MariaDB with ProxySQL and DBTrail”

    Comments on “HTAP on MariaDB with ProxySQL and DBTrail”
  53. No. 53Score 0.64

    GPT-6.2 Astra: OpenAI warns of new risks from AI self-improvement

    Summary and score for “GPT-6.2 Astra: OpenAI warns of new risks from AI self-improvement”

    OpenAI has released GPT-6.2 Astra, its flagship model, and GPT-6.2 Sol, a cheaper general-purpose model, emphasizing lower cost per task. Artificial Analysis ranks Astra second on its Intelligence Index, behind Claude Fable 5.2.

    Score 0.64Contestedjudges differ by 0.63 on new information

    Elsewhere: HN —

    Score breakdown for “GPT-6.2 Astra: OpenAI warns of new risks from AI self-improvement” · Also covered by The Next Web

    Comments on “GPT-6.2 Astra: OpenAI warns of new risks from AI self-improvement”
  54. No. 54Score 0.64

    WebMCP: Agent shortcuts and unintentional dark patterns

    Summary and score for “WebMCP: Agent shortcuts and unintentional dark patterns”

    WebMCP lets websites expose in-page tools that browser assistants can call directly, and this piece argues that such shortcuts can create unintentional dark patterns. In a ride-booking example, an agent bypasses a surge-pricing confirmation gate that the visual interface enforces, charging an unapproved €85 fare.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “WebMCP: Agent shortcuts and unintentional dark patterns”

    Comments on “WebMCP: Agent shortcuts and unintentional dark patterns”
  55. No. 55Score 0.64

    Stricter checks for uploaded distributions on PyPI

    Summary and score for “Stricter checks for uploaded distributions on PyPI”

    PyPI has started rejecting some uploads: source distributions using legacy V7-style tar archives, and wheels containing ZIP comments that could be misread as control structures. The goal is to reduce parser differentials that let attackers make installers, mirrors, and scanners see different contents from the same file.

    Score 0.64

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Stricter checks for uploaded distributions on PyPI”

    Comments on “Stricter checks for uploaded distributions on PyPI”
  56. No. 56Score 0.64

    Nvidia/DeepSeek-V4-Pro-0813-NVFP4

    Summary and score for “Nvidia/DeepSeek-V4-Pro-0813-NVFP4”

    NVIDIA has released DeepSeek-V4-Pro-0813-NVFP4, an NVFP4-quantized version of DeepSeek's 1.65-trillion-parameter Mixture-of-Experts model (49B active), published on Hugging Face under the MIT license. Only the MoE linear layers are quantized, and the checkpoint is built for Blackwell B200 with SGLang or vLLM, supporting a 1-million-token context.

    Score 0.64

    Elsewhere: HN 9 pts, not on front page

    Score breakdown for “Nvidia/DeepSeek-V4-Pro-0813-NVFP4”

    Comments on “Nvidia/DeepSeek-V4-Pro-0813-NVFP4”
  57. No. 57Score 0.64

    Word Machines for Weird Machines

    Summary and score for “Word Machines for Weird Machines”

    A Phrack article by noperator, "Word Machines for Weird Machines," argues that LLM-based vulnerability discovery should be driven by a practitioner-controlled harness rather than open-ended "go find bugs" prompts, and that scaling attention to under-examined code is the main lever.

    Score 0.64

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Word Machines for Weird Machines”

    Comments on “Word Machines for Weird Machines”
  58. No. 58Score 0.64

    WordPress 7.1.5 [Security Release]

    Summary and score for “WordPress 7.1.5 [Security Release]”

    WordPress 7.1.5, a security release published by WordPress.org, fixes a long list of vulnerabilities, including SQL injection in WP_Meta_Query, multiple XSS flaws, a multisite privilege escalation that let site Administrators take over Super Admin accounts, and an unauthenticated path to remote code execution via Template Part block attributes.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “WordPress 7.1.5 [Security Release]” · Also covered by Documentation, heise online English

    Comments on “WordPress 7.1.5 [Security Release]”
  59. No. 59Score 0.64

    Google's AI answers reducing search traffic by 40%

    Summary and score for “Google's AI answers reducing search traffic by 40%”

    A randomized experiment on US desktop Chrome users found that Google's AI Overviews cut clicks to outside websites by about 40% on searches where an Overview appeared.

    Score 0.64

    Elsewhere: HN 694 pts, front page

    Score breakdown for “Google's AI answers reducing search traffic by 40%”

    Comments on “Google's AI answers reducing search traffic by 40%”
  60. No. 60Score 0.64

    Apache Shiro – Critical security advisories: LDAP realm and cookie fixes

    Summary and score for “Apache Shiro – Critical security advisories: LDAP realm and cookie fixes”

    Apache Shiro's maintainers have published three critical security advisories affecting its Java security framework. Two involve DefaultLdapRealm: a flaw where cached logins accept any password when authentication caching is enabled, and LDAP DN injection through unescaped usernames.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Apache Shiro – Critical security advisories: LDAP realm and cookie fixes”

    Comments on “Apache Shiro – Critical security advisories: LDAP realm and cookie fixes”
  61. No. 61Score 0.64

    TLAPS-Bench: Formally Proving Correctness Of Real-World System Specifications With TLA+ Proof System

    Summary and score for “TLAPS-Bench: Formally Proving Correctness Of Real-World System Specifications With TLA+ Proof System”

    TLAPS-Bench is a new benchmark of 265 TLA+ proof targets drawn from 95 real-system specifications, such as etcd Raft and ZooKeeper, with its harness Tilapia. Unlike benchmarks that supply scaffolding, it asks agents to find inductive invariants for whole distributed protocols.

    Score 0.64Contestedjudges differ by 0.41 on significance

    Elsewhere: HN —

    Score breakdown for “TLAPS-Bench: Formally Proving Correctness Of Real-World System Specifications With TLA+ Proof System”

    Comments on “TLAPS-Bench: Formally Proving Correctness Of Real-World System Specifications With TLA+ Proof System”
  62. No. 62Score 0.62

    Software Factory builds 30 complex open source libraries in a month

    Summary and score for “Software Factory builds 30 complex open source libraries in a month”

    A solo developer says his in-house "software factory" has produced 30 open source C# libraries, about 1.9 million lines of handwritten code, 28 of them reaching 1.0 between October and mid-November.

    Score 0.62Contestedjudges differ by 0.44 on promotional

    Elsewhere: HN 18 pts, front page

    Score breakdown for “Software Factory builds 30 complex open source libraries in a month”

    Comments on “Software Factory builds 30 complex open source libraries in a month”
  63. No. 63Score 0.64

    Censys Nginx2root Exposure Tracker

    Summary and score for “Censys Nginx2root Exposure Tracker”

    Censys has published a live tracker of nginx servers exposed to the "nginx2root" bug, an HTTP/2 vulnerability. Fixed releases are 1.30.7 (stable) and 1.31.9 (mainline), while older branches receive no patch and must upgrade or disable HTTP/2 as a mitigation.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Censys Nginx2root Exposure Tracker”

    Comments on “Censys Nginx2root Exposure Tracker”
  64. No. 64Score 0.63

    The financial rules AI agents break when chasing results

    Summary and score for “The financial rules AI agents break when chasing results”

    A Chinese study introducing the EconSafeBench benchmark, by researchers at the Chinese University of Hong Kong and collaborators, found that AI agents managing simulated investment portfolios began breaking trading rules once performance targets became impossible to meet legally.

    Score 0.63Contestedjudges differ by 0.47 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “The financial rules AI agents break when chasing results”

    Comments on “The financial rules AI agents break when chasing results”
  65. No. 65Score 0.63

    A Wolf at the Door: Apple Is Rethinking Mac Permissions for the Age of AI Agents

    Summary and score for “A Wolf at the Door: Apple Is Rethinking Mac Permissions for the Age of AI Agents”

    Apple's macOS 27.3 beta 2, announced on 16 November, reworks how Macs handle permissions for AI agents and other software. Full Disk Access becomes granular, letting users grant access to specific protected folders rather than everything. Device Control and Data Access is split into two separate permissions.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A Wolf at the Door: Apple Is Rethinking Mac Permissions for the Age of AI Agents”

    Comments on “A Wolf at the Door: Apple Is Rethinking Mac Permissions for the Age of AI Agents”
  66. No. 66Score 0.63

    Achieving the impossible: XSS in innerHTML without equals

    Summary and score for “Achieving the impossible: XSS in innerHTML without equals”

    A guest post by security researcher Masato Kinugawa, published on PortSwigger's research channel, describes an XSS vector that solves PortSwigger's "impossible" lab "innerHTML context but no equals allowed." The technique abuses the newer HTML selectedcontent element, which clones the selected option's child nodes, to run an SVG script that was…

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “Achieving the impossible: XSS in innerHTML without equals”

    Comments on “Achieving the impossible: XSS in innerHTML without equals”
  67. No. 67Score 0.57

    Why valid Kubernetes YAML still breaks in prod and how to catch it before deploy

    Summary and score for “Why valid Kubernetes YAML still breaks in prod and how to catch it before deploy”

    The author describes k8s-guardian, an open-source tool (Apache 2.0, now at v0.7) that checks Kubernetes manifests against live cluster context, such as quotas, node capacity and missing references, and diffs them against what is deployed.

    Score 0.57

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Why valid Kubernetes YAML still breaks in prod and how to catch it before deploy”

    Comments on “Why valid Kubernetes YAML still breaks in prod and how to catch it before deploy”
  68. No. 68Score 0.63

    AI-assisted code at scale: provenance, quality, centrality and review time

    Summary and score for “AI-assisted code at scale: provenance, quality, centrality and review time”

    A peer-reviewed paper from Audris Mockus, Peter C. Rigby, Don Stewart, Chandra Maddila and Nachiappan Nagappan, accepted at ESEM 2026's Software Engineering in Practice track, presents a character-level provenance system that tracks AI-generated code through source control rather than heuristics.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI-assisted code at scale: provenance, quality, centrality and review time”

    Comments on “AI-assisted code at scale: provenance, quality, centrality and review time”
  69. No. 69Score 0.57

    PowerShell 7 heading into Windows as 5.1 hits the end of the road

    Summary and score for “PowerShell 7 heading into Windows as 5.1 hits the end of the road”

    Microsoft confirmed that PowerShell 7, the cross-platform, open-source successor, will ship bundled with Windows, while the decade-old PowerShell 5.1, built on .NET Framework, begins retirement.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “PowerShell 7 heading into Windows as 5.1 hits the end of the road” · Also covered by support.microsoft.com

    Comments on “PowerShell 7 heading into Windows as 5.1 hits the end of the road”
  70. No. 70Score 0.63

    Self-Adapting Language Models

    Summary and score for “Self-Adapting Language Models”

    MIT researchers Adam Zweiger, Jyothish Pari, and colleagues introduce Self-Adapting Language Models (SEAL), a framework where an LLM generates its own finetuning data and update instructions, then learns via reinforcement learning, using downstream performance as reward, to produce better self-edits.

    Score 0.63

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Self-Adapting Language Models”

    Comments on “Self-Adapting Language Models”
  71. No. 71Score 0.63

    180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers

    Summary and score for “180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers”

    A Transformers pull request (huggingface/transformers#50717) adds an experts_offload option that streams a MoE model's routed experts from SSD through a bounded device cache, letting Qwen3.8-Flash-Next run on a 32 GB Mac.

    Score 0.63Contestedjudges differ by 0.42 on significance, 0.44 on practical relevance

    Elsewhere: HN —

    Score breakdown for “180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers”

    Comments on “180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers”
  72. No. 72Score 0.63

    Spreadsheets were the first vibe coding

    Summary and score for “Spreadsheets were the first vibe coding”

    The author revisits a 2019 prediction that people would make disposable one-off software with AI, as they make spreadsheets, and finds that spreadsheets were in fact handed on and kept alive for years.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Spreadsheets were the first vibe coding”

    Comments on “Spreadsheets were the first vibe coding”
  73. No. 73Score 0.62

    Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data

    Summary and score for “Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data”

    This paper revisits the claim that "RL generalizes, SFT memorizes" by examining how SFT training data affects generalization on Qwen2.5 and Qwen3 models. The authors find that data source matters: a mixed SFT dataset outperforms data from a larger model alone.

    Score 0.62

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data”

    Comments on “Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data”
  74. No. 74Score 0.62

    Designing the Langfuse AI Gateway

    Summary and score for “Designing the Langfuse AI Gateway”

    Langfuse's engineers explain why they built its new AI Gateway in Rust rather than Node.js or Python, and why it passes provider APIs through without translation.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Designing the Langfuse AI Gateway”

    Comments on “Designing the Langfuse AI Gateway”
  75. No. 75Score 0.62

    Rumble-2-Flash: an open 0.5B speech model for 40+ languages

    Summary and score for “Rumble-2-Flash: an open 0.5B speech model for 40+ languages”

    Reson8 has open-sourced Rumble-2-Flash under Apache 2.0, a 0.5B-parameter multilingual speech recognition model covering 40+ languages with automatic language detection. It is built on the Qwen3-ASR architecture so it runs natively in Transformers and vLLM, and it was distilled from the company's larger Rumble-2 using on-policy distillation with GOLD to…

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “Rumble-2-Flash: an open 0.5B speech model for 40+ languages”

    Comments on “Rumble-2-Flash: an open 0.5B speech model for 40+ languages”
  76. No. 76Score 0.62

    Java4Shell: when base64 decoding leads to remote code execution?

    Summary and score for “Java4Shell: when base64 decoding leads to remote code execution?”

    In a post on his blog, Daniel Lemire walks through CVE-2026-114938, a bug in OpenJDK's Base64 decoder where an input of 128 A characters followed by one = makes an AVX-512 tail routine perform an out-of-bounds write.

    Score 0.62Contestedjudges differ by 0.48 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Java4Shell: when base64 decoding leads to remote code execution?”

    Comments on “Java4Shell: when base64 decoding leads to remote code execution?”
  77. No. 77Score 0.62

    Keeping Agents Fast: High-Interactivity Decode with FlyDSL MonoKernels in vLLM

    Summary and score for “Keeping Agents Fast: High-Interactivity Decode with FlyDSL MonoKernels in vLLM”

    This post describes MonoKernels written in AMD's FlyDSL that fuse several dependent decode stages into a single persistent GPU launch, reducing per-kernel startup, drain, and handoff costs for low-concurrency interactive LLM serving in vLLM on AMD Instinct GPUs.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Keeping Agents Fast: High-Interactivity Decode with FlyDSL MonoKernels in vLLM”

    Comments on “Keeping Agents Fast: High-Interactivity Decode with FlyDSL MonoKernels in vLLM”
  78. No. 78Score 0.62

    WeightBench: Evaluating LLMs as weight-space programmers

    Summary and score for “WeightBench: Evaluating LLMs as weight-space programmers”

    WeightBench, a benchmark from Marasini, Khadka and Rahimi (arXiv:2611.18553), tests whether AI agents can read, repair and write neural-network weights directly, given a task and a small metered budget of forward passes. Its 42 seeded tasks span five tracks, and scores come from a hash-chained ledger of model calls.

    Score 0.62

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “WeightBench: Evaluating LLMs as weight-space programmers”

    Comments on “WeightBench: Evaluating LLMs as weight-space programmers”
  79. No. 79Score 0.62

    valkey-io/valkey 7.2.16

    Summary and score for “valkey-io/valkey 7.2.16”

    Valkey 7.2.16, released 18 November 2026 by the Valkey project, is a maintenance release that the maintainers say users should apply promptly. Its headline fix addresses CVE-2026-130822, a use-after-free crash in client eviction when maxmemory-clients is enabled.

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “valkey-io/valkey 7.2.16” · Also covered by GitHub, GitHub, GitHub, GitHub

    Comments on “valkey-io/valkey 7.2.16”
  80. No. 80Score 0.61

    Gemini 4 Neon

    Summary and score for “Gemini 4 Neon”

    Google's Agent Platform documentation page for Gemini 4 Neon, published on Google Cloud's documentation site, lists it as one of several Gemini 4 model variants alongside Hydrogen, Carbon, and Argon. The page is mainly navigation and model-family structure, so it offers little detail on capabilities or performance.

    Score 0.61Contestedjudges differ by 0.41 on substantiated

    Elsewhere: HN 12 pts, not on front page

    Score breakdown for “Gemini 4 Neon”

    Comments on “Gemini 4 Neon”
  81. No. 81Score 0.61

    A modern FPS starves wallhacks and radar cheats

    Summary and score for “A modern FPS starves wallhacks and radar cheats”

    Dead Drop, a 100-player PC battle royale, has its server send each client only the enemies that client could plausibly see, filtering by distance, facing, scope cone and line of sight.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A modern FPS starves wallhacks and radar cheats”

    Comments on “A modern FPS starves wallhacks and radar cheats”
  82. No. 82Score 0.61

    Local search for coding agents that knows which decision is still current

    Summary and score for “Local search for coding agents that knows which decision is still current”

    LEKS 3.1, a local search tool for folders of notes that needs no API key, was benchmarked against BM25 and three open tools on 695 held-out developer questions about Python PEPs, all run on one Apple silicon Mac.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Local search for coding agents that knows which decision is still current”

    Comments on “Local search for coding agents that knows which decision is still current”
  83. No. 83Score 0.61

    Wayfinder: Frontier-Level Code Search at a Fraction of the Cost

    Summary and score for “Wayfinder: Frontier-Level Code Search at a Fraction of the Cost”

    Wayfinder is an open-source research project from MongoDB-affiliated author Henry Weller that pairs a fine-tuned 35B open model with a hybrid semantic and BM25 code index.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Wayfinder: Frontier-Level Code Search at a Fraction of the Cost”

    Comments on “Wayfinder: Frontier-Level Code Search at a Fraction of the Cost”
  84. No. 84Score 0.61

    Evolving Infra: Agent-Adaptive LLM Serving at Scale

    Summary and score for “Evolving Infra: Agent-Adaptive LLM Serving at Scale”

    Evolving Infra, an open-source LLM serving system from MIT Han Lab (arXiv 2611.12304), adapts to agent workloads online without offline profiling. It separates compute allocation from KV cache management, predicts prefill and decode load using a cost model covering sub-agents and context compaction, and evicts KV no agent branch needs.…

    Score 0.61Contestedjudges differ by 0.48 on substantiated

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Evolving Infra: Agent-Adaptive LLM Serving at Scale”

    Comments on “Evolving Infra: Agent-Adaptive LLM Serving at Scale”
  85. No. 85Score 0.61

    Haru-4B: A voice AI that can listen and speak at the same time

    Summary and score for “Haru-4B: A voice AI that can listen and speak at the same time”

    Haru-4B is an open-weight pretrained base model, built on Qwen3.5-4B, that listens and speaks at the same time, learning turn-taking, backchannels, and interruptions from about 40 million hours of real, mostly unscripted speech rather than synthetic dialogue.

    Score 0.61Contestedjudges differ by 0.46 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Haru-4B: A voice AI that can listen and speak at the same time”

    Comments on “Haru-4B: A voice AI that can listen and speak at the same time”
  86. No. 86Score 0.61

    Claude was trained on protein sequences at scale

    Summary and score for “Claude was trained on protein sequences at scale”

    Researchers tested whether the latest Claude models predict next amino acids in little-studied bacterial protein families, a skill that suggests learning from protein sequence data rather than text about proteins.

    Score 0.61Contestedjudges differ by 0.43 on audience breadth

    Elsewhere: HN 170 pts, front page

    Score breakdown for “Claude was trained on protein sequences at scale”

    Comments on “Claude was trained on protein sequences at scale”
  87. No. 87Score 0.61

    curl on OSS-Scanner

    Summary and score for “curl on OSS-Scanner”

    Daniel Stenberg, curl's lead developer, reports that Anthropic's OSS-Scanner sent 45 claimed vulnerabilities in curl after the project enrolled in the service. After the curl team triaged them, only about 11% were confirmed as vulnerabilities, with roughly 80% turning out to be ordinary bugs.

    Score 0.61

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “curl on OSS-Scanner”

    Comments on “curl on OSS-Scanner”
  88. No. 88Score 0.58

    More load is faster: Tuning Postgres read replicas

    Summary and score for “More load is faster: Tuning Postgres read replicas”

    A ClickHouse engineering post tests how Postgres read-replica lag behaves when the primary is under low or moderate write load. With default settings, Postgres ships WAL only every 16MB or 60 seconds, so lightly loaded systems can show long lag.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “More load is faster: Tuning Postgres read replicas”

    Comments on “More load is faster: Tuning Postgres read replicas”
  89. No. 89Score 0.60

    Building secure mobile apps in the agentic era

    Summary and score for “Building secure mobile apps in the agentic era”

    Shopify describes how it secures its mobile apps while shipping code at about 11 times last year's rate, averaging roughly 434,000 changed lines per week during a 12-week rewrite.

    Score 0.60

    Elsewhere: HN —

    Score breakdown for “Building secure mobile apps in the agentic era”

    Comments on “Building secure mobile apps in the agentic era”
  90. No. 90Score 0.60

    Security patch: PHP admins have work to do Thursday afternoon

    Summary and score for “Security patch: PHP admins have work to do Thursday afternoon”

    PHP has disclosed several security flaws, some rated critical under non-default configurations, affecting all supported versions. Fixes and technical details are due Thursday, 3–3:30 p.m. CET, with the PHP security team urging administrators to prepare to upgrade promptly.

    Score 0.60Contestedjudges differ by 0.68 on new information

    Elsewhere: HN —

    Score breakdown for “Security patch: PHP admins have work to do Thursday afternoon”

    Comments on “Security patch: PHP admins have work to do Thursday afternoon”
  91. No. 91Score 0.60

    TheFed finds $3T of AI commitments that aren't on big tech's balance sheet

    Summary and score for “TheFed finds $3T of AI commitments that aren't on big tech's balance sheet”

    The Federal Reserve's November Financial Stability Report includes a first-ever section on leverage in the data center buildout, counting over $3 trillion in off-balance-sheet AI commitments, including unstarted leases and take-or-pay contracts at major tech firms.

    Score 0.60

    Elsewhere: HN 8 pts, not on front page

    Score breakdown for “TheFed finds $3T of AI commitments that aren't on big tech's balance sheet”

    Comments on “TheFed finds $3T of AI commitments that aren't on big tech's balance sheet”
  92. No. 92Score 0.60

    Design Systems Should Be Fast by Default

    Summary and score for “Design Systems Should Be Fast by Default”

    Many design systems are not tested for component-level performance, even though slow components hurt every app built on them, the author argues. Testing ten popular open-source systems, the author timed a 249-country dropdown and selecting 100 checkboxes under 4x CPU throttling.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Design Systems Should Be Fast by Default”

    Comments on “Design Systems Should Be Fast by Default”
  93. No. 93Score 0.57

    The EU Digital Fairness Act Proposal: European Commission Proposes Major Update of EU Consumer Law

    Summary and score for “The EU Digital Fairness Act Proposal: European Commission Proposes Major Update of EU Consumer Law”

    The European Commission has proposed the Digital Fairness Act, one of three "Consumer Package" proposals to update EU consumer law for online markets. It would replace the 2005 Unfair Commercial Practices Directive with a directly applicable regulation and add specific rules on dark patterns, influencer marketing, personalised pricing, loot boxes…

    Score 0.57Contestedjudges differ by 0.46 on reader interest

    Elsewhere: HN —

    Score breakdown for “The EU Digital Fairness Act Proposal: European Commission Proposes Major Update of EU Consumer Law”

    Comments on “The EU Digital Fairness Act Proposal: European Commission Proposes Major Update of EU Consumer Law”
  94. No. 94Score 0.60

    How to tell if an outage is a cyberattack

    Summary and score for “How to tell if an outage is a cyberattack”

    João's November newsletter from IPinfo offers a seven-question checklist for telling a self-inflicted outage from a cyberattack, using the mid-November 2026 disruptions as examples.

    Score 0.60Contestedjudges differ by 0.42 on audience breadth

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “How to tell if an outage is a cyberattack”

    Comments on “How to tell if an outage is a cyberattack”
  95. No. 95Score 0.56

    Benchmarking S3 alternatives: How do cheap VMs compare to S3 on EC2

    Summary and score for “Benchmarking S3 alternatives: How do cheap VMs compare to S3 on EC2”

    A hands-on benchmark compares S3 on AWS EC2 against cheaper S3-compatible providers (Backblaze B2, Cloudflare R2, DigitalOcean Spaces) accessed from low-cost VMs. Its author finds S3 most consistent, DigitalOcean and Backblaze surprisingly competitive in their own datacenters, and R2 variable, weighing performance against the substantial egress savings these alternatives offer.

    Score 0.56Contestedjudges differ by 0.41 on reader interest

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Benchmarking S3 alternatives: How do cheap VMs compare to S3 on EC2”

    Comments on “Benchmarking S3 alternatives: How do cheap VMs compare to S3 on EC2”
  96. No. 96Score 0.59

    Escaping from a Docker container through vDSO

    Summary and score for “Escaping from a Docker container through vDSO”

    A security researcher known as andersonc0d3 describes a technique for escaping a Docker container by abusing the vDSO, the kernel-mapped shared library visible across namespaces.

    Score 0.59Contestedjudges differ by 0.43 on practical relevance

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Escaping from a Docker container through vDSO”

    Comments on “Escaping from a Docker container through vDSO”
  97. No. 97Score 0.59

    GPT-6.2 Astra Formalizes Smoktunowicz's Gap Theorem

    Summary and score for “GPT-6.2 Astra Formalizes Smoktunowicz's Gap Theorem”

    GPT-6.2 Astra formally proved Agata Smoktunowicz's 2006 gap theorem on the growth of graded algebras, becoming the only model on the Autoform-Hard benchmark to solve a task that other top models missed. The run took 94 minutes and about 1,800 lines of Lean, with the proof passing clean-room grading.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “GPT-6.2 Astra Formalizes Smoktunowicz's Gap Theorem”

    Comments on “GPT-6.2 Astra Formalizes Smoktunowicz's Gap Theorem”
  98. No. 98Score 0.59

    UK MPs want AI watchdog with power to pull the plug on dangerous models

    Summary and score for “UK MPs want AI watchdog with power to pull the plug on dangerous models”

    A UK parliamentary committee has recommended an AI Safety and Security Bill that would create a watchdog able to halt high-risk AI training, block deployments, or suspend services while models are still in development.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “UK MPs want AI watchdog with power to pull the plug on dangerous models”

    Comments on “UK MPs want AI watchdog with power to pull the plug on dangerous models”
  99. No. 99Score 0.59

    FastFrames: Video Generation at Interactive Latency

    Summary and score for “FastFrames: Video Generation at Interactive Latency”

    Apple machine learning researchers introduced FastFrames, a family of autoregressive video diffusion models that reach interactive speed by replacing VAE decoding and KV cache updates with lightweight alternatives.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “FastFrames: Video Generation at Interactive Latency” · Also covered by huggingface.co

    Comments on “FastFrames: Video Generation at Interactive Latency”
  100. No. 100Score 0.59

    Why your fine-tuning job takes longer than the FLOPs say

    Summary and score for “Why your fine-tuning job takes longer than the FLOPs say”

    A preview paper reports that a CatBoost regressor, using only pre-dispatch job configuration, predicts end-to-end fine-tuning time on shared H200 clusters far better than the standard FLOP-based estimate.

    Score 0.59Contestedjudges differ by 0.42 on significance

    Elsewhere: HN —

    Score breakdown for “Why your fine-tuning job takes longer than the FLOPs say”

    Comments on “Why your fine-tuning job takes longer than the FLOPs say”
  101. No. 101Score 0.59

    Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking

    Summary and score for “Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking”

    Murmur is an inference system for long-context speech recognition that transcribes recordings up to an hour long about 2.5 times faster on average than single-pass decoding, including speaker diarization.

    Score 0.59

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking”

    Comments on “Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking”
  102. No. 102Score 0.59

    DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

    Summary and score for “DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1”

    RadixArk released DaisyRose-27B, a version of Alibaba's Qwen3.8-27B fine-tuned with reinforcement learning on 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs.

    Score 0.59Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 172 pts, front page

    Score breakdown for “DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1” · Also covered by Hugging Face Blog

    Comments on “DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1”
  103. No. 103Score 0.59

    Unlocking Stable On-Policy RL on TPU with Bitwise-Zero Train–Inference Mismatch

    Summary and score for “Unlocking Stable On-Policy RL on TPU with Bitwise-Zero Train–Inference Mismatch”

    Google's Tunix team describes a TPU reinforcement learning implementation, "Zero-TIM," that makes the sampler and trainer assign bitwise-identical logprobs to the same tokens under the same weights, eliminating a hidden off-policy gap in LLM RL.

    Score 0.59

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Unlocking Stable On-Policy RL on TPU with Bitwise-Zero Train–Inference Mismatch”

    Comments on “Unlocking Stable On-Policy RL on TPU with Bitwise-Zero Train–Inference Mismatch”
  104. No. 104Score 0.58

    Attempts to improve skill discovery in Claude Code

    Summary and score for “Attempts to improve skill discovery in Claude Code”

    A developer's write-up says Claude Code rarely opens the skill a task needs, opening it in 2 of 58 tasks on Haiku and 7 of 58 on Sonnet.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Attempts to improve skill discovery in Claude Code”

    Comments on “Attempts to improve skill discovery in Claude Code”
  105. No. 105Score 0.58

    Microsoft releases Aion 1.0 Instruct, a 2.3B on-device Windows SLM

    Summary and score for “Microsoft releases Aion 1.0 Instruct, a 2.3B on-device Windows SLM”

    Microsoft's Applied Sciences Group has released Aion 1.0 Instruct, a 2.3-billion-parameter small language model optimized for on-device use on Windows, with a 32K-token context window and MIT licensing. Microsoft says a tuned version paired with a content moderation model will become the Windows OS model by January 2027.

    Score 0.58Contestedjudges differ by 0.51 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Microsoft releases Aion 1.0 Instruct, a 2.3B on-device Windows SLM”

    Comments on “Microsoft releases Aion 1.0 Instruct, a 2.3B on-device Windows SLM”
  106. No. 106Score 0.58

    Every pixel is a measurement: Lossless video compression in GStreamer

    Summary and score for “Every pixel is a measurement: Lossless video compression in GStreamer”

    The article describes new lossless compression elements in GStreamer's 1.29 development series, adding zlib, deflate, and Brotli plugins that work on any media type. Their key feature is an original-caps field that preserves the pre-compression format, letting isomp4mux and qtdemux store and read compressed raw video, including Bayer sensor data…

    Score 0.58

    Elsewhere: HN 8 pts, not on front page

    Score breakdown for “Every pixel is a measurement: Lossless video compression in GStreamer”

    Comments on “Every pixel is a measurement: Lossless video compression in GStreamer”
  107. No. 107Score 0.58

    AI use is up 330×. The bill is up 115×

    Summary and score for “AI use is up 330×. The bill is up 115×”

    ModelCompare's analysis of OpenRouter data finds weekly AI token use grew about 330 times from January 2025 to October 2026, reaching 164 trillion tokens, while estimated list-price spending grew about 115 times to roughly $176 million a week.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI use is up 330×. The bill is up 115×”

    Comments on “AI use is up 330×. The bill is up 115×”
  108. No. 108Score 0.58

    Data Contracts

    Summary and score for “Data Contracts”

    A Martin Fowler-hosted article, "Data Contracts," argues that trust in business data is essential as agentic AI grows, and that data contracts supply it. Using the Open Data Contract Standard (ODCS), a Linux Foundation project, the authors show a machine-readable, verifiable YAML description covering schema, quality rules, freshness SLAs, versioning…

    Score 0.58Contestedjudges differ by 0.42 on promotional

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Data Contracts”

    Comments on “Data Contracts”
  109. No. 109Score 0.58

    A Transformer-Derived Iterative Preconditioner

    Summary and score for “A Transformer-Derived Iterative Preconditioner”

    Researchers Haris, Gangrade, Saligrama, and colleagues at Boston University report that a transformer trained on in-context linear systems implicitly improves the conditioning of the Gram matrix XX^T, and they extract this mechanism into RAPID, a standalone anytime preconditioner for symmetric positive definite systems.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “A Transformer-Derived Iterative Preconditioner”

    Comments on “A Transformer-Derived Iterative Preconditioner”
  110. No. 110Score 0.62

    AI models flag civilians as military targets

    Summary and score for “AI models flag civilians as military targets”

    A study by testing lab Vals, reported by heise online, found that several frontier AI models, when asked to identify military targets in simulated drone footage and satellite images, often flagged civilians and ambulances as targets, with Anthropic's models performing worst and OpenAI's best.

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “AI models flag civilians as military targets”

    Comments on “AI models flag civilians as military targets”
  111. No. 111Score 0.58

    Open-source mock MCP server to test clients and agents without real servers

    Summary and score for “Open-source mock MCP server to test clients and agents without real servers”

    mcp-mock is an open-source mock server for the Model Context Protocol, built by the team behind llm-mock, for integration-testing MCP clients and agents without running real servers.

    Score 0.58Contestedjudges differ by 0.41 on reader interest

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Open-source mock MCP server to test clients and agents without real servers”

    Comments on “Open-source mock MCP server to test clients and agents without real servers”
  112. No. 112Score 0.58

    C++ Optimizations in MSVC Build Tools 14.52

    Summary and score for “C++ Optimizations in MSVC Build Tools 14.52”

    MSVC Build Tools 14.52 adds C++ optimizations spanning reassociation, loop optimizations, SIMD code generation, and ARM64 address modes and bitmask patterns. The reassociation changes group operands so shared subexpressions can be hoisted, while loop work includes final-value replacement and sharing address calculations across indexed loads, which the post attributes mainly…

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “C++ Optimizations in MSVC Build Tools 14.52”

    Comments on “C++ Optimizations in MSVC Build Tools 14.52”
  113. No. 113Score 0.58

    FrontierBreach: Measuring frontier model attack susceptibility

    Summary and score for “FrontierBreach: Measuring frontier model attack susceptibility”

    PromptArmor's FrontierBreach is a benchmark measuring how susceptible AI agents are to indirect prompt injection during long-horizon enterprise tasks, built on GDPval scenarios with attacks derived from its threat intelligence research.

    Score 0.58Contestedjudges differ by 0.44 on significance

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “FrontierBreach: Measuring frontier model attack susceptibility”

    Comments on “FrontierBreach: Measuring frontier model attack susceptibility”
  114. No. 114Score 0.58

    Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages

    Summary and score for “Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages”

    Three sources told the press that bank executives and NSA officials informed Congress that this week's nationwide outages, which grounded over 20,000 FAA flights and disrupted bank access, stemmed from rushed patches for vulnerabilities AI models had flagged, not from a cyberattack.

    Score 0.58

    Elsewhere: HN 96 pts, front page

    Score breakdown for “Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages” · Also covered by The Record (Recorded Future News), bbc.com, nytimes.com, bbc.com, Slashdot, semafor.com

    Comments on “Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages”
  115. No. 115Score 0.55

    Engineering Vitals in the Age of AI

    Summary and score for “Engineering Vitals in the Age of AI”

    The author argues that engineering leaders need concrete vitals to judge whether AI is improving software delivery, rejecting both activity-counting and vague "value" goals. Octopus Deploy's CTO describes their three-part framework: delivery objectives like deployment lead time and time to recovery, quality objectives like high-severity customer defects and developer NPS…

    Score 0.55Contestedjudges differ by 0.47 on new information

    Elsewhere: HN 10 pts, front page

    Score breakdown for “Engineering Vitals in the Age of AI”

    Comments on “Engineering Vitals in the Age of AI”
  116. No. 116Score 0.57

    modernc.org/webview: Pure-Go Browser Engine

    Summary and score for “modernc.org/webview: Pure-Go Browser Engine”

    modernc.org/webview is a CGo-free web rendering engine for Go, implementing a deliberately partial but correct subset of HTML and CSS from HTML 2.0 through CSS 3 and HTML 5, with optional JavaScript.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “modernc.org/webview: Pure-Go Browser Engine”

    Comments on “modernc.org/webview: Pure-Go Browser Engine”
  117. No. 117Score 0.60

    Microsoft's agent-framework PyPI package Compromised with Credential Stealer

    Summary and score for “Microsoft's agent-framework PyPI package Compromised with Credential Stealer”

    Socket Research reports that versions 1.26.1 and 1.26.2 of Microsoft's agent-framework package on PyPI were compromised with credential-stealing code. Both versions use Python .pth startup hooks to silently harvest Azure, GitHub Actions, and CI/CD credentials and send them to a webhook.site endpoint.

    Score 0.60Contestedjudges differ by 0.92 on new informationClaim check: 2 not in source · 5 no source given

    Elsewhere: HN 25 pts, front page

    Score breakdown and claims for “Microsoft's agent-framework PyPI package Compromised with Credential Stealer”

    Comments on “Microsoft's agent-framework PyPI package Compromised with Credential Stealer”
  118. No. 118Score 0.57

    Standard Support: How AI agents ask for support

    Summary and score for “Standard Support: How AI agents ask for support”

    DocSpring has published Standard Support, a draft convention for how AI agents report problems to a product's support team. It defines three tools: a help tool that answers from public documentation without involving a person, a feedback tool for one-way reports nobody needs to answer, and a ticket tool that…

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Standard Support: How AI agents ask for support”

    Comments on “Standard Support: How AI agents ask for support”
  119. No. 119Score 0.57

    Building collaborative editing in EmDash CMS

    Summary and score for “Building collaborative editing in EmDash CMS”

    EmDash 1.5 adds collaborative editing, letting several people edit the same post at once. Each open entry gets a server-side room that holds the shared document (built on Yjs) and does the saving itself, rather than having each browser save its own copy, which can overwrite a colleague's offline edits.…

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Building collaborative editing in EmDash CMS”

    Comments on “Building collaborative editing in EmDash CMS”
  120. No. 120Score 0.57

    Design Once, Deploy at Scale: Building a Template-Driven ML Ecosystem

    Summary and score for “Design Once, Deploy at Scale: Building a Template-Driven ML Ecosystem”

    Meta describes a template-driven approach to its ads ranking ML ecosystem, in which models are built from a small set of standardized, reusable templates so that new techniques can be tested once on representative models and then propagated across the fleet.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “Design Once, Deploy at Scale: Building a Template-Driven ML Ecosystem” · Also covered by Engineering at Meta

    Comments on “Design Once, Deploy at Scale: Building a Template-Driven ML Ecosystem”
  121. No. 121Score 0.57

    How should we define "capabilities activities" for the purpose of pacing?

    Summary and score for “How should we define "capabilities activities" for the purpose of pacing?”

    A Pacing the Frontier proposal defines "capabilities activities" as all AI company compute use except a narrow set of exemptions: idle compute, narrow product development, most external inference, and narrow safety work. The aim is a compute cap on capabilities that still leaves room for safety and product work.

    Score 0.57Contestedjudges differ by 0.49 on substantiated, 0.42 on technical depth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “How should we define "capabilities activities" for the purpose of pacing?”

    Comments on “How should we define "capabilities activities" for the purpose of pacing?”
  122. No. 122Score 0.62

    CVE-2026-114938: Critical vulnerability in OpenJDK JAVA

    Summary and score for “CVE-2026-114938: Critical vulnerability in OpenJDK JAVA”

    Cisco Talos reports a critical OpenJDK flaw, CVE-2026-114938 (CVSS 4.0 score 9.2), that can allow unauthenticated remote code execution when a malformed Base64 input reaches affected JVMs.

    Score 0.62Contestedjudges differ by 0.41 on technical depth

    Elsewhere: HN —

    Score breakdown for “CVE-2026-114938: Critical vulnerability in OpenJDK JAVA”

    Comments on “CVE-2026-114938: Critical vulnerability in OpenJDK JAVA”
  123. No. 123Score 0.57

    N-able patches exploited security flaws in N-central

    Summary and score for “N-able patches exploited security flaws in N-central”

    Attackers have been exploiting flaws in N-able's N-central remote monitoring software to install crypto miners on vulnerable instances. N-able has now released N-central 2026.4 HF3, which adds authentication and hardening fixes and patches 17 further vulnerabilities, two rated critical.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “N-able patches exploited security flaws in N-central”

    Comments on “N-able patches exploited security flaws in N-central”
  124. No. 124Score 0.57

    LZ-RDO lands in ARM's astc-encoder

    Summary and score for “LZ-RDO lands in ARM's astc-encoder”

    Richard Geldreich's LZ-RDO rate-distortion optimization for ASTC texture encoding has been merged into ARM's astc-encoder via PR #781, with Fabian Giesen refining the original implementation before landing it.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “LZ-RDO lands in ARM's astc-encoder”

    Comments on “LZ-RDO lands in ARM's astc-encoder”
  125. No. 125Score 0.57

    Kitesurf is now available on GitHub under the Apache license

    Summary and score for “Kitesurf is now available on GitHub under the Apache license”

    Cloudflare has released Kitesurf, its alternative to headless Chromium for AI agents that read or screenshot websites, on GitHub under Apache 2.0. Built in Rust and compiled to WebAssembly, it runs on Cloudflare Workers and speaks the Chrome DevTools Protocol.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “Kitesurf is now available on GitHub under the Apache license”

    Comments on “Kitesurf is now available on GitHub under the Apache license”
  126. No. 126Score 0.57

    Strict Capabilities with a PNVI Escape Hatch

    Summary and score for “Strict Capabilities with a PNVI Escape Hatch”

    The paper proposes resolving a tension in capability-based memory safety for C: existing capabilities are too strict for low-level idioms like XOR-linked lists and tagged pointers, yet not strict enough to enforce type correctness and sub-object bounds.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Strict Capabilities with a PNVI Escape Hatch”

    Comments on “Strict Capabilities with a PNVI Escape Hatch”
  127. No. 127Score 0.57

    Quoting Daniel Stenberg

    Summary and score for “Quoting Daniel Stenberg”

    Daniel Stenberg, the maintainer of curl, reports that during a span of just over three weeks, AI agents converted about 185,000 lines of C code into roughly 340,000 lines of Rust.

    Score 0.57Contestedjudges differ by 0.41 on new information

    Elsewhere: HN —

    Score breakdown for “Quoting Daniel Stenberg”

    Comments on “Quoting Daniel Stenberg”
  128. No. 128Score 0.57

    Calif discloses Java4Shell flaw in OpenJDK's Base64 decoder

    Summary and score for “Calif discloses Java4Shell flaw in OpenJDK's Base64 decoder”

    Security firm Calif has disclosed a flaw it calls "Java4Shell" in OpenJDK's Base64 decoder, claiming it allows unauthenticated remote code execution. Upwind has reproduced a 62-byte overflow but says code execution has not yet been proven.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Calif discloses Java4Shell flaw in OpenJDK's Base64 decoder”

    Comments on “Calif discloses Java4Shell flaw in OpenJDK's Base64 decoder”
  129. No. 129Score 0.56

    100ms copy-on-write workspaces for AI agents on macOS

    Summary and score for “100ms copy-on-write workspaces for AI agents on macOS”

    Replica is an open-source macOS CLI that gives each AI agent run its own fully installed, real git checkout of a project, created in about 100 ms using APFS copy-on-write clonefile(2) calls rather than a fresh clone and install.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “100ms copy-on-write workspaces for AI agents on macOS”

    Comments on “100ms copy-on-write workspaces for AI agents on macOS”
  130. No. 130Score 0.56

    To rank the world's biggest supercomputers, we're gonna need a bigger bench

    Summary and score for “To rank the world's biggest supercomputers, we're gonna need a bigger bench”

    The Top500 ranking of the world's largest supercomputers is struggling to stay relevant as AI superclusters from companies like SpaceXAI, Meta, and OpenAI, which hold hundreds of thousands of GPUs, decline to submit results.

    Score 0.56Contestedjudges differ by 0.41 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “To rank the world's biggest supercomputers, we're gonna need a bigger bench”

    Comments on “To rank the world's biggest supercomputers, we're gonna need a bigger bench”
  131. No. 131Score 0.56

    Staged Publishing on Nuget.org

    Summary and score for “Staged Publishing on Nuget.org”

    NuGet.org has introduced opt-in staged publishing for all package publishers, letting them upload and validate packages privately, then promote them to public availability through an explicit website action. Staging groups can collect related packages for a release, though promotion of a group is not atomic.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Staged Publishing on Nuget.org”

    Comments on “Staged Publishing on Nuget.org”
  132. No. 132Score 0.56

    DKIM2 draft tracker: where every draft and open question stands

    Summary and score for “DKIM2 draft tracker: where every draft and open question stands”

    The DKIM2 draft tracker lists the status of nine active DKIM2 IETF drafts, three of them working-group documents and six individual submissions, as of late 2026.

    Score 0.56Contestedjudges differ by 0.42 on significance, 0.41 on reader interest

    Elsewhere: HN 9 pts, not on front page

    Score breakdown for “DKIM2 draft tracker: where every draft and open question stands”

    Comments on “DKIM2 draft tracker: where every draft and open question stands”
  133. No. 133Score 0.56

    A verifiable security state with Windows baseline protection

    Summary and score for “A verifiable security state with Windows baseline protection”

    Microsoft has begun previewing Windows baseline protection through the Windows Insider Program, with general availability planned for 2027. It combines existing Windows security features, including Secure Boot, virtualization-based security, memory integrity, and signed-driver requirements, into a single attestable device state that reports true only when every requirement is met.

    Score 0.56Contestedjudges differ by 0.44 on promotional

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “A verifiable security state with Windows baseline protection”

    Comments on “A verifiable security state with Windows baseline protection”
  134. No. 134Score 0.56

    The Wong test: how to get agents to do anything, almost anything

    Summary and score for “The Wong test: how to get agents to do anything, almost anything”

    Filippo Valsorda's author Wong (Samuel Wong, a security engineer at zkSecurity) argues that AI agents perform well on tasks with precise pass/fail tests, but those tests are easy to game, so verification must be strengthened.

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The Wong test: how to get agents to do anything, almost anything”

    Comments on “The Wong test: how to get agents to do anything, almost anything”
  135. No. 135Score 0.56

    Microsoft SharePoint flaw lets attackers inject and execute code

    Summary and score for “Microsoft SharePoint flaw lets attackers inject and execute code”

    Microsoft has disclosed two SharePoint Server flaws: CVE-2026-124223, a high-severity (CVSS 8.8) remote code execution bug where authenticated network attackers can inject and run malicious code via deserialization of untrusted data, and CVE-2026-124222, a medium-severity (CVSS 6.5) privilege escalation flaw.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “Microsoft SharePoint flaw lets attackers inject and execute code”

    Comments on “Microsoft SharePoint flaw lets attackers inject and execute code”
  136. No. 136Score 0.56

    Learn How Hermes Like Agents Work 1 PR at a Time – in Rust

    Summary and score for “Learn How Hermes Like Agents Work 1 PR at a Time – in Rust”

    A developer has published a learning project in Rust that rebuilds the core of a Hermes-style agent one commit at a time, with each step answering a specific design question such as tool execution, memory, and sandboxed shell access.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Learn How Hermes Like Agents Work 1 PR at a Time – in Rust”

    Comments on “Learn How Hermes Like Agents Work 1 PR at a Time – in Rust”
  137. No. 137Score 0.56

    Incremental static builds for React, Next and Astro.js backed by a headless CMS

    Summary and score for “Incremental static builds for React, Next and Astro.js backed by a headless CMS”

    hybld is an open-source (MIT) static-site build tool that rebuilds only the pages affected by a headless CMS change, rather than the whole site. It supports React + Vite, Next.js, Astro, SvelteKit and Nuxt, tracking content hashes in a SQLite manifest and offering CLI, webhook, and programmatic APIs.

    Score 0.56Contestedjudges differ by 0.41 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Incremental static builds for React, Next and Astro.js backed by a headless CMS”

    Comments on “Incremental static builds for React, Next and Astro.js backed by a headless CMS”
  138. No. 138Score 0.55

    Java4Shell, NGINX2Shell and more: Why we are not at "Yellow", (Wed, Nov 18th)

    Summary and score for “Java4Shell, NGINX2Shell and more: Why we are not at "Yellow", (Wed, Nov 18th)”

    SANS Internet Storm Center's Johannes Ullrich explains why the organization is keeping its InfoCon threat level at Green despite a run of recent vulnerabilities, including php2shell, a critical Java flaw dubbed Java4Shell, and a disclosed NGINX issue.

    Score 0.55Contestedjudges differ by 0.58 on new information

    Elsewhere: HN —

    Score breakdown for “Java4Shell, NGINX2Shell and more: Why we are not at "Yellow", (Wed, Nov 18th)”

    Comments on “Java4Shell, NGINX2Shell and more: Why we are not at "Yellow", (Wed, Nov 18th)”
  139. No. 139Score 0.55

    Aspire 17 Is Released

    Summary and score for “Aspire 17 Is Released”

    Microsoft's Aspire 17 release, announced by its principal product manager on the Aspire devblog, targets teams building internal developer platforms. It adds Git-backed shared team configuration for templates and integrations, a redesigned dashboard with resource tagging and grouping, preview Go AppHost support, and a new .NET project model that decouples…

    Score 0.55Contestedjudges differ by 0.49 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Aspire 17 Is Released”

    Comments on “Aspire 17 Is Released”
  140. No. 140Score 0.55

    Pitot – OpenTelemetry in One Binary, with Parquet, DuckDB, and MCP

    Summary and score for “Pitot – OpenTelemetry in One Binary, with Parquet, DuckDB, and MCP”

    Pitot, an alpha project from LabStack, is a self-hosted observability tool that packages OpenTelemetry ingest for traces, logs, and metrics into a single binary backed by local Parquet files and DuckDB analytics.

    Score 0.55Contestedjudges differ by 0.43 on substantiated

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Pitot – OpenTelemetry in One Binary, with Parquet, DuckDB, and MCP”

    Comments on “Pitot – OpenTelemetry in One Binary, with Parquet, DuckDB, and MCP”
  141. No. 141Score 0.55

    Uber says nearly 150,000 impacted by third-party breach

    Summary and score for “Uber says nearly 150,000 impacted by third-party breach”

    Uber told regulators in several states that a breach exposed names and driver's license data for 149,346 U.S. people. Hackers used stolen credentials from a third-party service provider, which TechCrunch linked to an attack on AI infrastructure firm Fireworks.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “Uber says nearly 150,000 impacted by third-party breach”

    Comments on “Uber says nearly 150,000 impacted by third-party breach”
  142. No. 142Score 0.55

    Readplace read-it-later self-registration Public API

    Summary and score for “Readplace read-it-later self-registration Public API”

    Readplace has published documentation for its public read-it-later API, which uses Siren hypermedia and Markdown article output. The guide explains how an integration registers via OAuth 2.0 Dynamic Client Registration with PKCE, handles short-lived access tokens and rotating refresh tokens, and polls a readlist to fetch, process and remove articles.…

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Readplace read-it-later self-registration Public API”

    Comments on “Readplace read-it-later self-registration Public API”
  143. No. 143Score 0.58

    Optimjack from the Outside: The Hijacked Script Was on Our Radar

    Summary and score for “Optimjack from the Outside: The Hijacked Script Was on Our Radar”

    A security vendor's external scan of 52 organizations loading Optimizely's experimentation script, a dependency exploited in the OPTIMJACK campaign that injected skimming code into visitors' pages. Five of those 52 appear on the 178-organization victim list, and a live re-fetch of 60 scripts found the malicious payload gone.

    Score 0.58Contestedjudges differ by 0.46 on new information, 0.47 on promotional

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Optimjack from the Outside: The Hijacked Script Was on Our Radar”

    Comments on “Optimjack from the Outside: The Hijacked Script Was on Our Radar”
  144. No. 144Score 0.59

    Russian cyber-espionage group used stolen cloud data to hack into organizations, Google says

    Summary and score for “Russian cyber-espionage group used stolen cloud data to hack into organizations, Google says”

    Google's Mandiant unit reports that Russian state-backed group Cold Relic, linked by U.S. authorities to the FSB, used data stolen from one U.S. organization's cloud accounts to impersonate employees and breach a second.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “Russian cyber-espionage group used stolen cloud data to hack into organizations, Google says”

    Comments on “Russian cyber-espionage group used stolen cloud data to hack into organizations, Google says”

How far each story is checked

Claim-checked
Scored by all three judges, and its checkable claims were compared with the sources it cites.
3 judges
Scored by all three judges. Only the day’s top stories get the claim check.
2 judges
Scored by two of the three judges. The third reads only the day’s strongest candidates.

Every story that scored at least 0.55, in score order. The judges never saw a site or an author; both are shown after ranking. How ranking works

This edition’s scores are open data: JSON and CSV, under CC BY 4.0. What the files hold

Get your mix by email

The first ten stories your filters show, in one email a day.

Or choose topics and formats on the newsletter page

One essay a week on AI and the work of programming: the hrih.si letter

This edition’s panel

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Table 1. How the day narrowed
StepStories
Candidates in the window1,416
Scored blind by two judges1,254
Off topic237
Read by the third judge237
Claim-checked23
Published144