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 105 stories · 13 hidden by your filters

  1. No. 1Score 0.79

    Clojure: Deserialization of Untrusted Data Allows Remote Code Execution

    Summary and score for “Clojure: Deserialization of Untrusted Data Allows Remote Code Execution”

    A security advisory from the Clojure project, as published on GitHub, describes a remote code execution flaw in Clojure's deserialization path. Compiled Clojure functions implement java.util.Comparator, so a crafted serialized java.util.PriorityQueue can run attacker-supplied Clojure source during a single ObjectInputStream.readObject call, using only classes already in clojure.jar and the JDK.…

    Score 0.79Claim check: no checkable claims

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Clojure: Deserialization of Untrusted Data Allows Remote Code Execution”

    Comments on “Clojure: Deserialization of Untrusted Data Allows Remote Code Execution”
  2. No. 2Score 0.76

    PhpAfterShell: From Interrupted Session Writes to Remote Code Execution

    Summary and score for “PhpAfterShell: From Interrupted Session Writes to Remote Code Execution”

    depthfirst disclosed phpAfterShell (CVE-2026-126714, CVSS 4.0 score 9.2), a PHP flaw that bypasses the earlier php2shell fix. During multipart upload-progress updates, PHP rewrites the file-backed session in place with non-atomic pwrite() and ftruncate(). If the worker dies mid-update, the file can hold a valid but spliced session that later session_start() deserializes. Chained with a memory-corruption bug in a native PHP class, depthfirst reports remote code execution. Fixes shipped in 8.2.36, 8.3.37, 8.4.28, 8.5.13 and 8.6.0. Cloudflare says its WAF blocks known attack patterns.

    Score 0.76Claim check: no checkable claims

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “PhpAfterShell: From Interrupted Session Writes to Remote Code Execution” · Also covered by heise online English, Cloudflare Blog, Cloudflare Blog

    Comments on “PhpAfterShell: From Interrupted Session Writes to Remote Code Execution”
  3. No. 3Score 0.75

    Hostile Inference: Exploiting Docker Model Runner for Container Escapes

    Summary and score for “Hostile Inference: Exploiting Docker Model Runner for Container Escapes”

    Zero Day Initiative detailed how an unauthenticated argument-injection bug in Docker Model Runner (ZDI-26-150 / CVE-2026-28400) lets a default-privilege container pass flags to llama-server, which runs on the host. Those flags give file read and write, and by overwriting Docker Extension assets the researchers reached code execution on macOS and Windows. They also describe a macOS Seatbelt sandbox escape and a WSL2 VM-to-host pivot that Docker considers outside its threat model. The fix restricts flags to an allowlist. Several dates and version details were not verified.

    Score 0.75Claim check: 3 matches source · 5 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Hostile Inference: Exploiting Docker Model Runner for Container Escapes”

    Comments on “Hostile Inference: Exploiting Docker Model Runner for Container Escapes”
  4. No. 4Score 0.78

    Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.

    Summary and score for “Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.”

    HD Moore of Dragos sent nearly two million prompts through AI routers and inference resellers, each seeded with a unique fake secret such as AWS keys or private URLs, and tracked where they surfaced. Three providers appear to have been breached, with stolen keys used tens of thousands of times, and some resellers logged prompts. A Chinese lab's scraper fetched URLs from over half its prompts, including many marked Zero Data Retention. For builders, privacy flags and router defaults are not reliable; choose upstream providers deliberately.

    Score 0.78Contestedjudges differ by 0.82 on new informationClaim check: 8 matches source

    Elsewhere: HN 12 pts, front page

    Score breakdown and claims for “Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.”

    Comments on “Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.”
  5. No. 5Score 0.72

    NVIDIA/TensorRT-LLM v1.3.0

    Summary and score for “NVIDIA/TensorRT-LLM v1.3.0”

    NVIDIA's TensorRT-LLM v1.3.0 adds preview Rubin (SM107) support, new models such as DeepSeek-V4, Kimi K3 and Qwen3.5, and VisualGen for image and video generation served through trtllm-serve. It also removes the TensorRT backend, trtllm-build and the legacy C++ Executor, leaving PyTorch as the only backend, so existing engine-based deployments need migration. Telemetry is on by default, KV transfer timeouts now default to 60 seconds, and the TRITON MoE backend is deprecated. The release notes are the only source, and several claims remain unverified.

    Score 0.72Claim check: 1 matches source · 5 no source given · 2 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “NVIDIA/TensorRT-LLM v1.3.0”

    Comments on “NVIDIA/TensorRT-LLM v1.3.0”
  6. No. 6Score 0.71

    PHP 8.6 is ready and tightens session defaults

    Summary and score for “PHP 8.6 is ready and tightens session defaults”

    PHP 8.6.0 adds partial function application, a nanosecond-resolution Time\Duration type, an Io\Poll API for epoll and kqueue (not an async runtime), and clamp(). The change most likely to affect existing applications is session defaults: strict mode and HttpOnly are on, and SameSite is Lax unless php.ini overrides them. heise online says this can break cross-site POST flows such as SSO and payment redirects, mainly in Firefox and Safari, so test logins before upgrading. The reported security fixes and support dates were not independently verified.

    Score 0.71Contestedjudges differ by 0.48 on new informationClaim check: 2 matches source · 4 no source given · 2 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “PHP 8.6 is ready and tightens session defaults”

    Comments on “PHP 8.6 is ready and tightens session defaults”
  7. No. 7Score 0.72

    Building a Software Factory

    Summary and score for “Building a Software Factory”

    Runway, a fintech software company, describes handing each stage of software delivery to agents: writing to green CI, review, risk-based approval, browser QA, fixing main, and coordination. It reports that agents drove 38% of October merges, median first-approval wait fell from over two hours to about five minutes, and AI-approved merges were reverted about one in a thousand against one in two hundred for human-approved ones. The figures are self-reported from internal data without external verification, and the mix of low-risk changes skews the revert comparison.

    Score 0.72Claim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Building a Software Factory”

    Comments on “Building a Software Factory”
  8. No. 8Score 0.70

    TensorLean: Formal verification framework for GPU kernels at the PTX level

    Summary and score for “TensorLean: Formal verification framework for GPU kernels at the PTX level”

    TensorLean, from researchers at Yale-affiliated authors Wei Qiang, Ganxiang Yang, Ronghui Gu and colleagues, is a Lean-based framework that verifies GPU kernels at the PTX assembly level for all input shapes and thread schedules, within a floating-point tolerance. The authors report checking 149 open-source kernel instances, including 102 from FlashAttention, proving 83 correct and finding 38 bugs. They also report a roughly 15:1 code-to-specification ratio. These figures are self-reported and uncorroborated, and the paper and code are not yet released.

    Score 0.70Claim check: 2 matches source · 5 no source given · 1 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “TensorLean: Formal verification framework for GPU kernels at the PTX level”

    Comments on “TensorLean: Formal verification framework for GPU kernels at the PTX level”
  9. No. 9Score 0.70

    Making microvms fast, persistence and cheap on GKE with Kata and CH

    Summary and score for “Making microvms fast, persistence and cheap on GKE with Kata and CH”

    Evgeniy Pats of kosem describes how the company runs coding-agent environments as Kata Containers microVMs on Cloud Hypervisor on GKE. Most of the speedup came from storage and memory accounting, not the hypervisor. Using raw block ext4 instead of virtiofs cut a Linux source unpack from about 126 s to under 9 s, and cold boot now takes 20 s. Lowering Kata's default guest memory fixed host OOM kills. The figures are kosem's own measurements, from two runs on one node type.

    Score 0.70Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Making microvms fast, persistence and cheap on GKE with Kata and CH”

    Comments on “Making microvms fast, persistence and cheap on GKE with Kata and CH”
  10. No. 10Score 0.69

    Google will launch Android app verification requirement globally in January 2027

    Summary and score for “Google will launch Android app verification requirement globally in January 2027”

    Google says Android developer verification will start rolling out in all regions in January 2027, reaching all devices within "several months." After a September launch in Brazil, Indonesia, Singapore and Thailand, sideloaded apps will require a developer registered with Google, using government ID and a $25 fee. Google reports "millions" of non-Play apps registered, covering over 99 percent of installs, and partner stores including Galaxy Store and Xiaomi will enforce it. Ars Technica notes developers have objected; the security benefit is Google's claim.

    Score 0.69Claim check: 4 matches source · 2 no source given · 2 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Google will launch Android app verification requirement globally in January 2027”

    Comments on “Google will launch Android app verification requirement globally in January 2027”
  11. No. 11Score 0.73

    GnuPG Security Advisory for Gpgme (T8557,CVE-2026-129975)

    Summary and score for “GnuPG Security Advisory for Gpgme (T8557,CVE-2026-129975)”

    GnuPG released GPGME 2.3.1 to fix a double free that a crafted OpenPGP key with a very long user ID can trigger. Werner Koch's advisory says it could leak application data and that remote code execution can't be ruled out. The cause is a reallocated buffer pointer saved to the context only at function end, leaving a stale pointer on early return. Software using earlier versions, including statically linked builds, is affected. Rebuild or patch; Gpg4win 5.1.2 ships the fix. These details come only from the advisory.

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

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “GnuPG Security Advisory for Gpgme (T8557,CVE-2026-129975)”

    Comments on “GnuPG Security Advisory for Gpgme (T8557,CVE-2026-129975)”
  12. No. 12Score 0.72

    Microsoft 365 outage of 12 and 13 November 2026

    Summary and score for “Microsoft 365 outage of 12 and 13 November 2026”

    Microsoft's preliminary post-incident report, summarised by 365Fanatic, attributes the 12–13 November Microsoft 365 outage, about 24 hours with Exchange Online hit hardest, to a latent defect in a Windows fix rolled out to its networking gateway servers in July. A Teams failover raised load while nearly 14% of capacity was offline, and the defect cascaded until over 80% of servers failed. Months-old changes blocked quick rollback, and test labs lacked high-core hardware. The report is preliminary; repair items are due 20 November.

    Score 0.72Claim check: 8 matches source

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Microsoft 365 outage of 12 and 13 November 2026”

    Comments on “Microsoft 365 outage of 12 and 13 November 2026”
  13. No. 13Score 0.68

    Valkey 9.2 offers forkless snapshots, compressed replication, ACL roles and more

    Summary and score for “Valkey 9.2 offers forkless snapshots, compressed replication, ACL roles and more”

    Valkey 9.2 adds opt-in forkless RDB snapshots and full syncs, which run on a background thread to avoid copy-on-write memory growth, at 4 extra bytes per key. Replication can be LZ4-compressed; a pull-request benchmark showed 52% less bandwidth with under 1% throughput cost. Large sorted sets move to a B+ tree, with a reported 36–43% lower per-member overhead (not independently verified). The release also adds ACL roles, a Path Hash type, conditional commands and hot key detection. The project advises checking downgrade compatibility first.

    Score 0.68Claim check: 5 matches source · 1 no source given · 2 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Valkey 9.2 offers forkless snapshots, compressed replication, ACL roles and more”

    Comments on “Valkey 9.2 offers forkless snapshots, compressed replication, ACL roles and more”
  14. No. 14Score 0.69

    AI is giving everyone the same advice

    Summary and score for “AI is giving everyone the same advice”

    Developer Cody Collier queried 65 models from 30 companies about 128,000 times and found heavily converged answers: 73% recommended the same book, 90% said to study computer science, and 69% of invented people were aged 34. Answers also identify the maker, with 59% of models closest to a sibling versus 4% by chance. In OLMo, the habit appears at supervised fine-tuning. Listing ten options and picking by number, or giving personal details, reduced overlap. The figures are Collier's own, with no independent source.

    Score 0.69Contestedjudges differ by 0.44 on substantiatedClaim check: 8 no source given

    Elsewhere: HN 70 pts, front page

    Score breakdown and claims for “AI is giving everyone the same advice”

    Comments on “AI is giving everyone the same advice”
  15. No. 15Score 0.67

    Pyrefly v1.5: 100% Conformance and 40% Faster

    Summary and score for “Pyrefly v1.5: 100% Conformance and 40% Faster”

    Pyrefly v1.5, the Python type checker from the Pyrefly team, reports passing all 146 files in the typing conformance suite and running about 40% faster than v1.4.1 (geometric mean of wall time across 71 projects, with 9% less memory). Against ty it is roughly even: faster on pytorch and sympy, slower on most smaller projects. It also adds a configurable hint severity. These figures are the team's own measurements, taken on one M5 Max laptop, and could not be independently verified.

    Score 0.67Claim check: 5 no source given · 3 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Pyrefly v1.5: 100% Conformance and 40% Faster”

    Comments on “Pyrefly v1.5: 100% Conformance and 40% Faster”
  16. No. 16Score 0.74

    nginx2root: Five new nginx vulnerabilities, and how Cloudflare protects origins from the worst of them

    Summary and score for “nginx2root: Five new nginx vulnerabilities, and how Cloudflare protects origins from the worst of them”

    F5 and nginx.org disclosed ten nginx flaws on November 18, five of them the memory-corruption bugs Xint named nginx2root. The worst, CVE-2026-123007, affects nginx with HTTP/2 enabled since 1.9.5. Cloudflare says origins reachable only through its proxy are shielded from the HTTP/2 and HTTP/3 bugs, but not the mail proxy bug and only partly the resolver bug. Google Cloud says its core load balancing is unaffected. Fixes are in nginx 1.30.7 and 1.31.9. The severity ratings and disclosure count were not verified.

    Score 0.74Contestedjudges differ by 0.79 on new information, 0.69 on promotionalClaim check: 1 matches source · 1 no source given · 1 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “nginx2root: Five new nginx vulnerabilities, and how Cloudflare protects origins from the worst of them” · Also covered by Google Cloud Documentation

    Comments on “nginx2root: Five new nginx vulnerabilities, and how Cloudflare protects origins from the worst of them”
  17. No. 17Score 0.65

    When Cache Hits Starve the Event Loop

    Summary and score for “When Cache Hits Starve the Event Loop”

    Octri's Emad Khair shows that awaiting two million in-memory cache hits can starve an event loop. A 10 ms timer fired after about 270 ms in Python and 142 ms in Node.js, because neither runtime handed control back during the loop. Yielding every 1,024 lookups, with asyncio.sleep(0) or setImmediate, let the timer fire near 10 ms, but the batch finished later (336 versus 270 ms in Python, 163 versus 142 ms in Node). The test used five runs per variant and one timer, and no production traffic.

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

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “When Cache Hits Starve the Event Loop”

    Comments on “When Cache Hits Starve the Event Loop”
  18. No. 18Score 0.69

    Inkling 2 Mini

    Summary and score for “Inkling 2 Mini”

    Thinking Machines Lab released Inkling 2 Mini, an open-weights looped transformer: a 26.4B-parameter mixture-of-experts model that runs its middle layers three times per token. The ~20GB footprint fits a laptop, and vLLM supports it from day one. The company reports the top open-weights scores under 100B parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%); these figures are self-reported, and setups vary by model. It trails on multi-turn chat and constraint-following. A variable-loop checkpoint supports research on looping and chain-of-thought monitorability.

    Score 0.69Claim check: 8 no source given

    Elsewhere: HN 850 pts, front page

    Score breakdown and claims for “Inkling 2 Mini” · Also covered by vLLM Blog

    Comments on “Inkling 2 Mini”
  19. No. 19Score 0.72

    Abusing Browser-Embedded Language Models for Post-Exploitation

    Summary and score for “Abusing Browser-Embedded Language Models for Post-Exploitation”

    Researchers at the Army Cyber Institute and NVIDIA show that Chrome's built-in Gemini Nano can be driven from the host using only PowerShell or Python, with a synthetic click satisfying the download gesture check. In one simulated operation, reading a password from a webcam image and staging an SSH key, it succeeded in 41 of 50 trials given an explicit plan, versus 60% with an open-ended goal. Google marked the reports Won't Fix. Results come from one laptop and one scenario, so defenders should review enterprise policies.

    Score 0.72Claim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Abusing Browser-Embedded Language Models for Post-Exploitation”

    Comments on “Abusing Browser-Embedded Language Models for Post-Exploitation”
  20. No. 20Score 0.71

    Brake Check: Hijacking Trailer Brakes over Bluetooth

    Summary and score for “Brake Check: Hijacking Trailer Brakes over Bluetooth”

    Element Digital researchers Patrick Forck and Jeremy McCarthy report that Elecbrakes trailer brake controllers accepted a four-byte Bluetooth write that applied the brakes at any force, with no pairing or authentication. Their testing used their own hardware. CISA's advisory lists three CVEs, the headline one rated CVSS 9.6. A firmware fix for the EB2 and VM1 shipped in July, but the older ELBC 2000 has none. The firmware update channel also uses CRC checks, not signatures. Several figures come only from the researchers and Elecbrakes.

    Score 0.71Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Brake Check: Hijacking Trailer Brakes over Bluetooth”

    Comments on “Brake Check: Hijacking Trailer Brakes over Bluetooth”
  21. No. 21Score 0.65

    Jujutsu Primer

    Summary and score for “Jujutsu Primer”

    Jujutsu Primer is a practical introduction to Jujutsu (jj), a version control system that works alongside existing Git repositories, written for someone already fluent in Git.

    Score 0.65Claim check: 1 matches source · 5 no source given · 2 couldn’t check

    Elsewhere: HN 198 pts, front page

    Score breakdown and claims for “Jujutsu Primer” · Also covered by brajeshwar.com

    Comments on “Jujutsu Primer”
  22. No. 22Score 0.62

    Squeezing Tables with Postgres 19

    Summary and score for “Squeezing Tables with Postgres 19”

    Postgres 19 ships REPACK, which folds VACUUM FULL and CLUSTER into one command and adds a CONCURRENTLY option that compacts tables without blocking reads or writes. Written by Robert Haas, it builds on Antonin Houska's pg_squeeze, using logical decoding of WAL to replay changes made during the copy.

    Score 0.62Claim check: 1 matches source · 3 no source given · 4 couldn’t check

    Elsewhere: HN 5 pts, not on front page

    Score breakdown and claims for “Squeezing Tables with Postgres 19”

    Comments on “Squeezing Tables with Postgres 19”
  23. No. 23Score 0.68

    Mapping every neuron in Qwen3-8B

    Summary and score for “Mapping every neuron in Qwen3-8B”

    Researchers mapped individual neurons in Qwen3-8B and found that pushing a single one, at every word, can switch the model's persona or language: one neuron turns it into a pirate in 100% of 30 test questions, while 30 random neurons from the same layer produce no such change.

    Score 0.68

    Elsewhere: HN 36 pts, front page

    Score breakdown for “Mapping every neuron in Qwen3-8B”

    Comments on “Mapping every neuron in Qwen3-8B”
  24. No. 24Score 0.70

    China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server

    Summary and score for “China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server”

    Amazon CISO CJ Moses reports that suspected China-linked attackers used the open-source Cairn penetration-testing framework to run AI agents that autonomously discovered and exploited a zero-day in MONAI Label, a medical imaging application, to seize control of a university server.

    Score 0.70Claim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server” · Also covered by The Record (Recorded Future News)

    Comments on “China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server”
  25. No. 25Score 0.61

    JEDEC Previews DDR6 Memory Standard: 2.35x More Bandwidth, 44% Narrower DIMMs

    Summary and score for “JEDEC Previews DDR6 Memory Standard: 2.35x More Bandwidth, 44% Narrower DIMMs”

    JEDEC has previewed its in-development DDR6 memory standard, outlining per-pin speeds starting at 12.8 GT/s and reaching 14.4 GT/s or higher, plus a move to a 96-bit DIMM bus and a more compact 74mm module.

    Score 0.61

    Elsewhere: HN —

    Score breakdown for “JEDEC Previews DDR6 Memory Standard: 2.35x More Bandwidth, 44% Narrower DIMMs”

    Comments on “JEDEC Previews DDR6 Memory Standard: 2.35x More Bandwidth, 44% Narrower DIMMs”
  26. No. 26Score 0.61

    UX Roundup: GPT-6.2 | Prepare with AI | Speed | Unambitious Academics | AI Permission Requests | Permissions Barrage | Recognition-Primed Decisions | Directed Editing | Misleading Defaults

    Summary and score for “UX Roundup: GPT-6.2 | Prepare with AI | Speed | Unambitious Academics | AI Permission Requests | Permissions Barrage | Recognition-Primed Decisions | Directed Editing | Misleading Defaults”

    A UX roundup piece argues that OpenAI's new GPT-6.2 Astra, despite its benchmark lead, produces weaker agentic video work than Anthropic's Fable 5.2 and even the cheaper Opus 5.5, based on the author's own side-by-side videos and viewing statistics.

    Score 0.61Contestedjudges differ by 0.41 on audience breadth, 0.42 on overclaim

    Elsewhere: HN —

    Score breakdown for “UX Roundup: GPT-6.2 | Prepare with AI | Speed | Unambitious Academics | AI Permission Requests | Permissions Barrage | Recognition-Primed Decisions | Directed Editing | Misleading Defaults”

    Comments on “UX Roundup: GPT-6.2 | Prepare with AI | Speed | Unambitious Academics | AI Permission Requests | Permissions Barrage | Recognition-Primed Decisions | Directed Editing | Misleading Defaults”
  27. No. 27Score 0.61

    Xinproof – Restore-test your database backups, get a signed receipt

    Summary and score for “Xinproof – Restore-test your database backups, get a signed receipt”

    xinproof is a pre-release (0.2.0) open-source CLI that restore-tests database backups for PostgreSQL, MySQL/MariaDB and MongoDB. It checks integrity, restores each dump into a fresh, resource-limited Docker container, compares per-table row counts and index health against the dump, and signs an Ed25519 receipt for each result, hash-chained to the previous…

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Xinproof – Restore-test your database backups, get a signed receipt”

    Comments on “Xinproof – Restore-test your database backups, get a signed receipt”
  28. No. 28Score 0.69

    OpenSSL: Buffer overflow in EVP_EncodeBlock() (CVE-2026-121481, CVSS 9.8)

    Summary and score for “OpenSSL: Buffer overflow in EVP_EncodeBlock() (CVE-2026-121481, CVSS 9.8)”

    The OpenSSL project's pull request addresses CVE-2026-121481, a CVSS 9.8 buffer overflow in the EVP_EncodeBlock() base64 encoding function. The change adds input guards that return an error for lengths whose encoded output would overflow, and extends similar checks to EVP_DecodeBlock() and to base64encode() in http_client.c.

    Score 0.69

    Elsewhere: HN 365 pts, front page

    Score breakdown for “OpenSSL: Buffer overflow in EVP_EncodeBlock() (CVE-2026-121481, CVSS 9.8)” · Also covered by cve.now

    Comments on “OpenSSL: Buffer overflow in EVP_EncodeBlock() (CVE-2026-121481, CVSS 9.8)”
  29. No. 29Score 0.69

    VMware patches actively exploited security flaws

    Summary and score for “VMware patches actively exploited security flaws”

    Broadcom has released patches for a critical authentication-bypass flaw in VMware vCenter (CVE-2026-123890, CVSS 9.8), which Mandiant reports is being actively exploited by a group it links to China, alongside several high and medium flaws in SDDC Manager and VMware Tools.

    Score 0.69

    Elsewhere: HN —

    Score breakdown for “VMware patches actively exploited security flaws”

    Comments on “VMware patches actively exploited security flaws”
  30. No. 30Score 0.68

    Attackers hid M365 takeovers with inbox rules named "a", "cc" and " "

    Summary and score for “Attackers hid M365 takeovers with inbox rules named "a", "cc" and " "”

    Researchers at Reco describe a takeover of eight Microsoft 365 accounts at one organization in late 2026, likely via adversary-in-the-middle phishing and stolen sessions. Each takeover followed the same pattern: an inbox rule with a 1–3 character name and no conditions, deleting or hiding incoming mail, appearing within minutes.

    Score 0.68

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Attackers hid M365 takeovers with inbox rules named "a", "cc" and " "”

    Comments on “Attackers hid M365 takeovers with inbox rules named "a", "cc" and " "”
  31. No. 31Score 0.60

    Building the Infrastructure Underneath Next.js

    Summary and score for “Building the Infrastructure Underneath Next.js”

    Easel, a managed hosting platform, reports that its infrastructure now passes 937 of 940 Next.js deployment test files that run in deployment mode, up from 639 in its first complete run against Next.js 16.5.0-canary.48.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Building the Infrastructure Underneath Next.js”

    Comments on “Building the Infrastructure Underneath Next.js”
  32. No. 32Score 0.68

    Language Models Can Rewrite Their Own Weights

    Summary and score for “Language Models Can Rewrite Their Own Weights”

    A 27B open-weight language model can rewrite its own weights through its own reasoning, with no gradients or training, and the authors report that the ability emerges with scale. In a paired test on 490 unseen problems, its self-edit raised accuracy by 6.6 points.

    Score 0.68

    Elsewhere: HN 373 pts, front page

    Score breakdown for “Language Models Can Rewrite Their Own Weights”

    Comments on “Language Models Can Rewrite Their Own Weights”
  33. No. 33Score 0.67

    Merkle Tree Certificates

    Summary and score for “Merkle Tree Certificates”

    The IETF PLANTS working group has published draft-ietf-plants-merkle-tree-certs-08, a standards-track Internet-Draft from Google, Apple, Cloudflare and Geomys defining Merkle Tree certificates. These are X.509 certificates that integrate public Certificate Transparency-style logging directly into the certificate.

    Score 0.67

    Elsewhere: HN 29 pts, front page

    Score breakdown for “Merkle Tree Certificates”

    Comments on “Merkle Tree Certificates”
  34. No. 34Score 0.59

    Nostromo – Hosted Kubernetes control planes that sleep when idle

    Summary and score for “Nostromo – Hosted Kubernetes control planes that sleep when idle”

    Nostromo is an open-source program, posted as a Show HN project, that hosts many unmodified Kubernetes control planes in one Go binary and suspends each one when idle: asleep, standby (API server only), or full.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Nostromo – Hosted Kubernetes control planes that sleep when idle”

    Comments on “Nostromo – Hosted Kubernetes control planes that sleep when idle”
  35. No. 35Score 0.59

    How deep is the Riemann hypothesis? An interactive explainer

    Summary and score for “How deep is the Riemann hypothesis? An interactive explainer”

    A Hacker News "Show HN" post presents an interactive explainer of the Riemann hypothesis, the unproven claim that all nontrivial zeros of the zeta function have real part one-half.

    Score 0.59Contestedjudges differ by 0.51 on reader interest, 0.42 on overclaim

    Elsewhere: HN 9 pts, not on front page

    Score breakdown for “How deep is the Riemann hypothesis? An interactive explainer”

    Comments on “How deep is the Riemann hypothesis? An interactive explainer”
  36. No. 36Score 0.65

    Postgres 19 is out, and it's available on Neon and Databricks Lakebase

    Summary and score for “Postgres 19 is out, and it's available on Neon and Databricks Lakebase”

    Databricks announced that PostgreSQL 19 is available on its Lakebase managed service on the same day it reached general availability, rather than waiting weeks as is typical for managed providers.

    Score 0.65Contestedjudges differ by 0.70 on new information, 0.58 on promotional

    Elsewhere: HN 63 pts, front page

    Score breakdown for “Postgres 19 is out, and it's available on Neon and Databricks Lakebase” · Also covered by Certuvia, heise online English

    Comments on “Postgres 19 is out, and it's available on Neon and Databricks Lakebase”
  37. No. 37Score 0.58

    Vibecoding an Inference Chip

    Summary and score for “Vibecoding an Inference Chip”

    A blog post by an engineer recounts vibe-coding an FPGA to run a ternary-weight LLM on AWS F2 instances, using a coding agent over several days for about $10,000 in tokens.

    Score 0.58

    Elsewhere: HN 132 pts, front page

    Score breakdown for “Vibecoding an Inference Chip”

    Comments on “Vibecoding an Inference Chip”
  38. No. 38Score 0.58

    Written Agent Authorizations: 25 Instruments, Dissected for AI-Agent Design

    Summary and score for “Written Agent Authorizations: 25 Instruments, Dissected for AI-Agent Design”

    This 2026 paper analyzes 25 government and private authorization instruments, including California's CCPA authorized-agent rules, to derive design patterns for AI-agent delegation. Its central claim is that most hard problems in agent authorization, such as scoping, revocation, counterparty refusal and sub-delegation, already have tested answers in existing forms, usually in…

    Score 0.58Contestedjudges differ by 0.52 on significance, 0.60 on reader interest, 0.47 on audience breadth, 0.53 on technical depth, 0.46 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Written Agent Authorizations: 25 Instruments, Dissected for AI-Agent Design”

    Comments on “Written Agent Authorizations: 25 Instruments, Dissected for AI-Agent Design”
  39. No. 39Score 0.58

    A TPM that BitLocker trusts, on Apple's Virtualization framework

    Summary and score for “A TPM that BitLocker trusts, on Apple's Virtualization framework”

    MacVisor 27.2.1 gives Windows 11 VMs on Apple's Virtualization framework a TPM 2.0 that BitLocker can seal to, despite the framework having no TPM device.

    Score 0.58

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “A TPM that BitLocker trusts, on Apple's Virtualization framework” · Also covered by GitHub

    Comments on “A TPM that BitLocker trusts, on Apple's Virtualization framework”
  40. No. 40Score 0.66

    Combatting adversarial distillation

    Summary and score for “Combatting adversarial distillation”

    OpenAI describes how it detects and disrupts adversarial distillation, where bad actors train a model on another model's outputs to copy its capabilities without its safeguards.

    Score 0.66

    Elsewhere: HN 34 pts, front page

    Score breakdown for “Combatting adversarial distillation” · Also covered by 霧雲

    Comments on “Combatting adversarial distillation”
  41. No. 41Score 0.65

    Papr-embed-code SOTA at code retrieval open source

    Summary and score for “Papr-embed-code SOTA at code retrieval open source”

    Papr's papr-embed-code-1 is an open-weight code embedding model, built by further training Qwen3-Embedding on code retrieval, in 0.6B and 4B sizes. On code tasks it never trained on, both sizes beat their base model by about 0.03 nDCG@10, while prose tasks scored slightly lower.

    Score 0.65Contestedjudges differ by 0.51 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Papr-embed-code SOTA at code retrieval open source”

    Comments on “Papr-embed-code SOTA at code retrieval open source”
  42. No. 42Score 0.61

    PowerShell 7 is coming to Windows

    Summary and score for “PowerShell 7 is coming to Windows”

    Microsoft's PowerShell team announced that PowerShell 7 will ship with Windows as the upgrade to Windows PowerShell 5.1, which has long run much of the world's Windows automation. Existing scripts need no changes, since the two versions run side by side.

    Score 0.61Contestedjudges differ by 0.42 on reader interest

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “PowerShell 7 is coming to Windows”

    Comments on “PowerShell 7 is coming to Windows”
  43. No. 43Score 0.59

    ChronoStack-move time forward for your Docker Compose stack

    Summary and score for “ChronoStack-move time forward for your Docker Compose stack”

    ChronoStack is an alpha open-source tool, published on GitHub by Asmodeus14, that runs a copy of a Docker Compose stack on a shared virtual clock. It moves Postgres, Redis and application services forward together, fires declared scheduled jobs at their virtual times, and checks what each service sees.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “ChronoStack-move time forward for your Docker Compose stack”

    Comments on “ChronoStack-move time forward for your Docker Compose stack”
  44. No. 44Score 0.66

    Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost

    Summary and score for “Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost”

    The post presents a method for training small LoRA "harness adapters" for Qwen3.8-27B using reinforcement learning inside each agent harness (Pi, OpenCode, Hermes Agent) on a custom coding task set.

    Score 0.66

    Elsewhere: HN —

    Score breakdown for “Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost”

    Comments on “Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost”
  45. No. 45Score 0.56

    California subpoenas Nvidia over whether chips or secrets reached China

    Summary and score for “California subpoenas Nvidia over whether chips or secrets reached China”

    California Attorney General Rob Bonta has served Nvidia with an investigative subpoena asking whether its advanced chips or confidential information reached China and whether the company misled investors.

    Score 0.56Contestedjudges differ by 0.42 on audience breadth

    Elsewhere: HN —

    Score breakdown for “California subpoenas Nvidia over whether chips or secrets reached China” · Also covered by Slashdot

    Comments on “California subpoenas Nvidia over whether chips or secrets reached China”
  46. No. 46Score 0.64

    AI is building me a web browser

    Summary and score for “AI is building me a web browser”

    An independent developer reports that Claude, working autonomously with no technical input and no paid API usage, has spent about seven weeks building SWB, a from-scratch web browser in Rust of roughly 400,000 lines of code.

    Score 0.64Contestedjudges differ by 0.43 on significance, 0.42 on substantiated

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “AI is building me a web browser”

    Comments on “AI is building me a web browser”
  47. No. 47Score 0.56

    AI boom threatens circuit board crunch to rival pandemic-era chip shortage

    Summary and score for “AI boom threatens circuit board crunch to rival pandemic-era chip shortage”

    A Global Electronics Association report warns of a printed circuit board shortage that could rival the pandemic-era chip crunch, driven by AI server demand.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “AI boom threatens circuit board crunch to rival pandemic-era chip shortage”

    Comments on “AI boom threatens circuit board crunch to rival pandemic-era chip shortage”
  48. No. 48Score 0.64

    Top AI models have mostly stopped inventing citations, Stanford study

    Summary and score for “Top AI models have mostly stopped inventing citations, Stanford study”

    A Stanford study by Owen Queen and James Zou, not yet peer reviewed, re-tested 50 widely cited findings about LLM weaknesses on 20 newer models and found that 44 per cent had expired, with the best current models almost never inventing citations.

    Score 0.64Contestedjudges differ by 0.45 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Top AI models have mostly stopped inventing citations, Stanford study”

    Comments on “Top AI models have mostly stopped inventing citations, Stanford study”
  49. No. 49Score 0.60

    Announcing DuckDB 2.0.1 LTS

    Summary and score for “Announcing DuckDB 2.0.1 LTS”

    DuckDB 2.0.1, the first patch release of its 2.0 LTS line, ships bugfixes, performance work, and security patches one month after 2.0.0. Notable changes include stricter SECRET scope matching, restoring per-match UPDATE ...

    Score 0.60

    Elsewhere: HN —

    Score breakdown for “Announcing DuckDB 2.0.1 LTS” · Also covered by GitHub

    Comments on “Announcing DuckDB 2.0.1 LTS”
  50. No. 50Score 0.56

    Windows Server 2028 trades compatibility for out-of-the-box security

    Summary and score for “Windows Server 2028 trades compatibility for out-of-the-box security”

    Microsoft's Windows Server Team, in a blog post, says a Windows Server 2028 preview tightens security defaults out of the box, which may break software relying on older protocols, legacy authentication, or scripts. Kerberos is now preferred over NTLM, while NTLM blocking and hardware-backed protections come in later previews.

    Score 0.56

    Elsewhere: HN 11 pts, not on front page

    Score breakdown for “Windows Server 2028 trades compatibility for out-of-the-box security”

    Comments on “Windows Server 2028 trades compatibility for out-of-the-box security”
  51. No. 51Score 0.64

    GitLab 19.6 adds AI agent audit log, but NGINX remains vulnerable

    Summary and score for “GitLab 19.6 adds AI agent audit log, but NGINX remains vulnerable”

    GitLab 19.6 adds audit logging for AI agent actions through its MCP server, plus server-side enforcement of tool blocks (still beta and flag-gated). Its bundled NGINX remains exposed to five nginx2root flaws, including an unauthenticated HTTP/2 bug, since the fixes exist only in NGINX 1.30.7 and 1.31.9.

    Score 0.64Contestedjudges differ by 0.51 on new information

    Elsewhere: HN —

    Score breakdown for “GitLab 19.6 adds AI agent audit log, but NGINX remains vulnerable”

    Comments on “GitLab 19.6 adds AI agent audit log, but NGINX remains vulnerable”
  52. No. 52Score 0.64

    GPT-6.2 Sol and Astra review: should you switch?

    Summary and score for “GPT-6.2 Sol and Astra review: should you switch?”

    CodeRabbit's evaluation of OpenAI's GPT-6.2 Sol and Astra models finds modest recall gains on code review benchmarks, but with substantially more comments and, in Astra's case, lower precision.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “GPT-6.2 Sol and Astra review: should you switch?”

    Comments on “GPT-6.2 Sol and Astra review: should you switch?”
  53. No. 53Score 0.63

    Crazy Agent Scale: Companies All-in on AI

    Summary and score for “Crazy Agent Scale: Companies All-in on AI”

    At a Sutter Hill Ventures event, engineers from several AI tooling companies described running agents at scale. Clement Pang outlined how his open-source Intent system coordinates agents through shared specs, task graphs, and on-demand waking to cut token costs.

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “Crazy Agent Scale: Companies All-in on AI”

    Comments on “Crazy Agent Scale: Companies All-in on AI”
  54. No. 54Score 0.63

    GnuPG's GPGME Affected By Critical Security Bug In Turn Affecting Mail Clients, Flatpak

    Summary and score for “GnuPG's GPGME Affected By Critical Security Bug In Turn Affecting Mail Clients, Flatpak”

    A critical security bug in GnuPG's GPGME library, the Made Easy cryptography interface, was announced by Werner Koch on a Friday afternoon. Because GPGME is widely embedded, the flaw reportedly reaches most email clients, most XMPP clients, Flatpak, and Samba, making it a broad supply-chain concern for product teams.

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “GnuPG's GPGME Affected By Critical Security Bug In Turn Affecting Mail Clients, Flatpak”

    Comments on “GnuPG's GPGME Affected By Critical Security Bug In Turn Affecting Mail Clients, Flatpak”
  55. No. 55Score 0.64

    AI incidents of 2026: Rogue agents, stolen keys, and other familiar stories

    Summary and score for “AI incidents of 2026: Rogue agents, stolen keys, and other familiar stories”

    A review of 50 AI-related security incidents from 2026 found that most trace back to familiar failures: exposed credentials, over-broad permissions, weak isolation, and compromised vendors. AI mainly changed the speed and reach of attacks, and agents sometimes acted beyond their intended limits.

    Score 0.64Contestedjudges differ by 0.49 on new information, 0.68 on promotional

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “AI incidents of 2026: Rogue agents, stolen keys, and other familiar stories”

    Comments on “AI incidents of 2026: Rogue agents, stolen keys, and other familiar stories”
  56. No. 56Score 0.68

    Ruby2Shell

    Summary and score for “Ruby2Shell”

    Aikido published a security advisory on Ruby2Shell (CVE-2026-132825), a critical one-byte heap overflow in Ruby's case-conversion code for legacy single-byte encodings such as ISO-8859-1, present since Ruby 2.4.0. It can let an unauthenticated attacker run code on affected servers, exposing default Rails and classic Sinatra apps.

    Score 0.68Contestedjudges differ by 0.71 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Ruby2Shell”

    Comments on “Ruby2Shell”
  57. No. 57Score 0.62

    Uncovering 100+ errors in AI’s most popular benchmarks

    Summary and score for “Uncovering 100+ errors in AI’s most popular benchmarks”

    Beaker, an autonomous dataset-review tool, flagged more than 100 errors in task instructions and grading across five widely used AI benchmarks, verifying at least one error in each.

    Score 0.62Contestedjudges differ by 0.45 on promotional, 0.42 on overclaim

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Uncovering 100+ errors in AI’s most popular benchmarks”

    Comments on “Uncovering 100+ errors in AI’s most popular benchmarks”
  58. No. 58Score 0.62

    OWASP Top:2025

    Summary and score for “OWASP Top:2025”

    OWASP has released the 2025 edition of its Top 10, a consensus awareness document listing the most critical security risks for web applications. This version reflects updated data and trends, and adds new categories such as software supply chain failures and mishandling of exceptional conditions.

    Score 0.62

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “OWASP Top:2025”

    Comments on “OWASP Top:2025”
  59. No. 59Score 0.56

    Trump admin still won't let Wi-Fi giant TP-Link sell its latest routers in US

    Summary and score for “Trump admin still won't let Wi-Fi giant TP-Link sell its latest routers in US”

    The Trump administration's March 2026 ban on new foreign-made consumer routers still blocks TP-Link, a leading US supplier, from selling its newest Wi-Fi 8 models in the US, while competitors like Netgear, Asus, and Starlink have won waivers.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Trump admin still won't let Wi-Fi giant TP-Link sell its latest routers in US” · Also covered by Slashdot

    Comments on “Trump admin still won't let Wi-Fi giant TP-Link sell its latest routers in US”
  60. No. 60Score 0.62

    Crafting Targeted MP4 Fuzzers

    Summary and score for “Crafting Targeted MP4 Fuzzers”

    A security research post describes why conventional mutational fuzzers struggle with MP4's nested, size-dependent atom structure, and how the author built a baseline fuzzer for GStreamer's qtdemux element.

    Score 0.62

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Crafting Targeted MP4 Fuzzers”

    Comments on “Crafting Targeted MP4 Fuzzers”
  61. No. 61Score 0.60

    Cutting my Zsh startup time: count the processes, not the lines

    Summary and score for “Cutting my Zsh startup time: count the processes, not the lines”

    A detailed guide on cutting Zsh startup time from around 500ms to 238ms on an older Mac, and from 213ms to 136ms on an M4, by profiling the config, switching from Oh My Zsh to Antidote, caching slow tool initializations, and above all cutting down the number of processes spawned…

    Score 0.60Contestedjudges differ by 0.43 on substantiated

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Cutting my Zsh startup time: count the processes, not the lines”

    Comments on “Cutting my Zsh startup time: count the processes, not the lines”
  62. No. 62Score 0.62

    Small TTS Bench: open-weight text-to-speech models on CPU

    Summary and score for “Small TTS Bench: open-weight text-to-speech models on CPU”

    A CPU benchmark of 21 small open-weight text-to-speech models measured first-audio latency, real-time speed, peak memory, and intelligibility (Whisper large-v3 word error rate) on a single thread.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Small TTS Bench: open-weight text-to-speech models on CPU”

    Comments on “Small TTS Bench: open-weight text-to-speech models on CPU”
  63. No. 63Score 0.62

    ReasoningBank: Enabling agents to learn from experience (April 2026)

    Summary and score for “ReasoningBank: Enabling agents to learn from experience (April 2026)”

    Google Research's ReasoningBank is a memory framework that lets LLM agents distill reusable reasoning strategies from both successful and failed runs, rather than storing raw action logs or only successful workflows.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “ReasoningBank: Enabling agents to learn from experience (April 2026)”

    Comments on “ReasoningBank: Enabling agents to learn from experience (April 2026)”
  64. No. 64Score 0.56

    Intel 18A yields hit 70% as company signs Microsoft as EMIB customer

    Summary and score for “Intel 18A yields hit 70% as company signs Microsoft as EMIB customer”

    Intel CEO Lip-Bu Tan said on Thursday that yields on the company's 18A (1.8nm-class) process have reached 70%, a level he described as roughly in line with expectations for a new manufacturing technology at this stage of ramp.

    Score 0.56

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “Intel 18A yields hit 70% as company signs Microsoft as EMIB customer”

    Comments on “Intel 18A yields hit 70% as company signs Microsoft as EMIB customer”
  65. No. 65Score 0.61

    Drone Hacking Part 2: Authorized ~_~ by Potensic ~_

    Summary and score for “Drone Hacking Part 2: Authorized ~_~ by Potensic ~_”

    A security researcher's second post in a series on the Potensic Atom 2 drone reverse-engineers the drone's firmware, app, and remote-control protocols. Findings so far include a WiFi access point password derived from the publicly broadcast device ID, which lets anyone in range reach stored photos and videos once Bluetooth…

    Score 0.61

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Drone Hacking Part 2: Authorized ~_~ by Potensic ~_”

    Comments on “Drone Hacking Part 2: Authorized ~_~ by Potensic ~_”
  66. No. 66Score 0.61

    Linux/EasyShell, a new Rust RAT implant and beacon aims Linux and macOS

    Summary and score for “Linux/EasyShell, a new Rust RAT implant and beacon aims Linux and macOS”

    A Chinese-language threat analysis describes "Linux/EasyShell," a Rust-based remote access implant delivered by a shell stager that targets Linux and macOS, served from a single C2 host at 192.3.31.148.

    Score 0.61

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Linux/EasyShell, a new Rust RAT implant and beacon aims Linux and macOS”

    Comments on “Linux/EasyShell, a new Rust RAT implant and beacon aims Linux and macOS”
  67. No. 67Score 0.60

    Microsoft's outage should terrify every enterprise – InfoWorld

    Summary and score for “Microsoft's outage should terrify every enterprise – InfoWorld”

    InfoWorld columnist argues that Microsoft's recent multi-day outage, which disrupted Microsoft 365, Teams, Outlook, Exchange Online and Copilot across more than 70 countries, exposes a deeper architectural problem rather than a simple bad update.

    Score 0.60

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Microsoft's outage should terrify every enterprise – InfoWorld”

    Comments on “Microsoft's outage should terrify every enterprise – InfoWorld”
  68. No. 68Score 0.59

    Our S3 cache is full of holes – We put them there

    Summary and score for “Our S3 cache is full of holes – We put them there”

    The post explains how Berserk, a log, metrics, trace and eval data store, uses sparse files on Linux as an S3 cache. Query nodes store only needed 4 MiB blocks, punch holes with fallocate to reclaim disk during eviction, and coordinate with query workers through a shared presence bitmap and…

    Score 0.59

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Our S3 cache is full of holes – We put them there”

    Comments on “Our S3 cache is full of holes – We put them there”
  69. No. 69Score 0.61

    Signal Is Serious About Protecting Your Privacy

    Summary and score for “Signal Is Serious About Protecting Your Privacy”

    Signal Technology Foundation says it is blocking screenshots in its app on iOS 27 and on Android when AI assistants try to read the screen, to stop OS-level AI agents from capturing private messages.

    Score 0.61

    Elsewhere: HN 537 pts, front page

    Score breakdown for “Signal Is Serious About Protecting Your Privacy” · Also covered by The Verge, The Next Web, The Next Web, reuters.com, The Register, Slashdot

    Comments on “Signal Is Serious About Protecting Your Privacy”
  70. No. 70Score 0.61

    Jeveration: System One Models Now Generate

    Summary and score for “Jeveration: System One Models Now Generate”

    Jeveration proposes that System One models, which answer by fast parallel prediction, can generate full images in a single forward pass rather than through many sequential steps.

    Score 0.61

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Jeveration: System One Models Now Generate” · Also covered by carol-lyh.github.io

    Comments on “Jeveration: System One Models Now Generate”
  71. No. 71Score 0.61

    Depth Pro 2 ONNX/TensorRT and ROS 2 (278 FPS on 3090)

    Summary and score for “Depth Pro 2 ONNX/TensorRT and ROS 2 (278 FPS on 3090)”

    An independent developer, Christian Engel, published unofficial ONNX and TensorRT exports of Apple's Depth Pro 2 monocular depth model, in focal-free (caller supplies focal length) and focal-estimating variants, plus ROS 2 and Jetson support.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Depth Pro 2 ONNX/TensorRT and ROS 2 (278 FPS on 3090)”

    Comments on “Depth Pro 2 ONNX/TensorRT and ROS 2 (278 FPS on 3090)”
  72. No. 72Score 0.60

    Debian APT to drop legacy features, citing advances in AI security research

    Summary and score for “Debian APT to drop legacy features, citing advances in AI security research”

    Debian developer Julian Andres Klode announced that APT will drop several legacy repository features from unstable starting the week of November 24, 2026, including detached Release.gpg signatures, weak-hash-only repositories, and repositories without hashes or a Release file.

    Score 0.60Contestedjudges differ by 0.55 on substantiated, 0.62 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Debian APT to drop legacy features, citing advances in AI security research”

    Comments on “Debian APT to drop legacy features, citing advances in AI security research”
  73. No. 73Score 0.60

    Agent Guard – limit what an AI agent can do when it gets tricked

    Summary and score for “Agent Guard – limit what an AI agent can do when it gets tricked”

    Agent Guard is an Apache-2.0 TypeScript library (version 1.4.0, Node.js 22+) that sits between an AI agent and its tools, allowing only calls that a task was explicitly granted and holding destructive actions for per-action human approval.

    Score 0.60Contestedjudges differ by 0.44 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Agent Guard – limit what an AI agent can do when it gets tricked”

    Comments on “Agent Guard – limit what an AI agent can do when it gets tricked”
  74. No. 74Score 0.60

    I vibecoded 70k LOC with DeepSeek and let Opus review it

    Summary and score for “I vibecoded 70k LOC with DeepSeek and let Opus review it”

    A developer built a self-hosted Apple Photos replacement (about 70k lines, Rust and SvelteKit) largely by prompting DeepSeek Flash 4.1, without reading the code, then had Claude Opus 5.5 run a 29-axis review.

    Score 0.60Contestedjudges differ by 0.44 on substantiated

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “I vibecoded 70k LOC with DeepSeek and let Opus review it”

    Comments on “I vibecoded 70k LOC with DeepSeek and let Opus review it”
  75. No. 75Score 0.60

    OpenAI Codex local log audit: 516-token cliff and zero-reasoning drops

    Summary and score for “OpenAI Codex local log audit: 516-token cliff and zero-reasoning drops”

    An open-source, offline, read-only forensic audit tool for local OpenAI Codex logs, SQLite databases, and installed binaries, written in Node.js and Python, claims to show that high-tier reasoning requests frequently return zero reasoning tokens and hit a 516-token cutoff even when subscription quota is under 80%.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “OpenAI Codex local log audit: 516-token cliff and zero-reasoning drops”

    Comments on “OpenAI Codex local log audit: 516-token cliff and zero-reasoning drops”
  76. No. 76Score 0.60

    Agent should not take the cursor

    Summary and score for “Agent should not take the cursor”

    A developer describes building a headless-browser computer-use agent that runs in its own Chrome instance so the user can keep working during long tasks. The main lesson concerns "walls" like captchas, passwords and passkeys: their wall detector silently missed passkey prompts because passkey screens lack a password field, so the…

    Score 0.60Contestedjudges differ by 0.43 on substantiated

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Agent should not take the cursor”

    Comments on “Agent should not take the cursor”
  77. No. 77Score 0.60

    Eighteen Things a Passkey Can Do

    Summary and score for “Eighteen Things a Passkey Can Do”

    A guide to eighteen uses for synced passkeys beyond the login button, covering passwordless sign-in, passwordless account creation, second-factor use, cross-device sign-in, app-to-website sharing, and shared credentials across company domains. Each entry names its trade-off and cites FIDO, NIST, Apple, Google, and Microsoft documentation.

    Score 0.60Contestedjudges differ by 0.46 on new informationClaim check: 3 matches source · 1 not in source · 4 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Eighteen Things a Passkey Can Do”

    Comments on “Eighteen Things a Passkey Can Do”
  78. No. 78Score 0.60

    Longshore: A 36 GB/S Zero-Allocation I/O and Object Storage Layer in Rust

    Summary and score for “Longshore: A 36 GB/S Zero-Allocation I/O and Object Storage Layer in Rust”

    Longshore is a roughly 25,000-line Rust I/O and object storage layer built on one io_uring ring per core, with no async runtime, no data-path locks, and all heap allocation done at startup.

    Score 0.60

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “Longshore: A 36 GB/S Zero-Allocation I/O and Object Storage Layer in Rust”

    Comments on “Longshore: A 36 GB/S Zero-Allocation I/O and Object Storage Layer in Rust”
  79. No. 79Score 0.59

    DigiCert: Unauthorized Access to Business System

    Summary and score for “DigiCert: Unauthorized Access to Business System”

    DigiCert, a certificate authority, self-reported on October 20, 2026 that an internal business system holding customer contact and identity-verification records was exposed externally after a February 2025 infrastructure migration. An account-registration feature lacking multi-factor authentication let unauthorized read-only accounts access archived data between June 27 and July 2, 2026.

    Score 0.59Contestedjudges differ by 0.48 on audience breadth

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “DigiCert: Unauthorized Access to Business System”

    Comments on “DigiCert: Unauthorized Access to Business System”
  80. No. 80Score 0.59

    HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning

    Summary and score for “HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning”

    HyperThink proposes replacing long chain-of-thought traces in large language models with a single query-conditioned weight update: a lightweight hypernetwork reads the question once and predicts a quantized bias update to a small subset of the frozen model's parameters, after which the model answers concisely.

    Score 0.59Contestedjudges differ by 0.48 on reader interest

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning”

    Comments on “HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning”
  81. No. 81Score 0.59

    Cursor tells lawmakers it's considering phasing out Chinese AI models

    Summary and score for “Cursor tells lawmakers it's considering phasing out Chinese AI models”

    CNBC reports that Cursor, the AI coding platform acquired by SpaceX in August, told House committees it is exploring phasing out Chinese open-weight models because of geopolitical risk, after building its Composer 2 model on Moonshot AI's Kimi K2.5.

    Score 0.59

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Cursor tells lawmakers it's considering phasing out Chinese AI models”

    Comments on “Cursor tells lawmakers it's considering phasing out Chinese AI models”
  82. No. 82Score 0.59

    Mind the Gap

    Summary and score for “Mind the Gap”

    Geoffrey Challen's essay "Mind the Gap" argues that a widening divide exists between how much people understand AI can do and what it can actually do, and that computing education reflects this.

    Score 0.59Contestedjudges differ by 0.48 on significance, 0.40 on substantiated, 0.58 on technical depth

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Mind the Gap”

    Comments on “Mind the Gap”
  83. No. 83Score 0.59

    AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure

    Summary and score for “AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure”

    METR's report finds that AI R&D-relevant capabilities of public models, including Opus 5.5 and GPT-6 Astra, are improving rapidly while becoming harder to measure. Models usually match or beat skilled humans on coding and optimization tasks, and time horizons on a combined suite are doubling roughly every 3.6 months.

    Score 0.59Claim check: 1 not in source · 7 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure” · Also covered by MIT Technology Review

    Comments on “AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure”
  84. No. 84Score 0.59

    Scala Without the JVM: Close to Rust's Speed, with 25x Less Memory Than the JVM

    Summary and score for “Scala Without the JVM: Close to Rust's Speed, with 25x Less Memory Than the JVM”

    Uni's HTTP server, built with Scala Native, reportedly reaches about 80% of Rust-with-axum throughput on a 3.4 KB response and about 80% of its own JVM build on that same response, while using roughly 25x less memory than the JVM (27 MB versus 650 MB).

    Score 0.59

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Scala Without the JVM: Close to Rust's Speed, with 25x Less Memory Than the JVM”

    Comments on “Scala Without the JVM: Close to Rust's Speed, with 25x Less Memory Than the JVM”
  85. No. 85Score 0.59

    Aura: 1.3B speech-to-video model that runs in real time on a single H100 GPU

    Summary and score for “Aura: 1.3B speech-to-video model that runs in real time on a single H100 GPU”

    Kyutai, with Sorbonne Université and Google DeepMind researchers, released AURA-1.3B, a 1.3B-parameter model that turns a portrait image and streaming speech audio into a talking-head video in real time at 576p.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Aura: 1.3B speech-to-video model that runs in real time on a single H100 GPU”

    Comments on “Aura: 1.3B speech-to-video model that runs in real time on a single H100 GPU”
  86. No. 86Score 0.58

    Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth

    Summary and score for “Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth”

    Apple researchers introduce Depth Pro 2, a single network that estimates metric depth in meters at full resolution for photos, videos, and live camera feeds. Prior methods traded temporal consistency against boundary sharpness, and this model aims to deliver both, with code and weights released publicly.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth” · Also covered by huggingface.co

    Comments on “Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth”
  87. No. 87Score 0.58

    Petsem: Sometimes Dead Is Better

    Summary and score for “Petsem: Sometimes Dead Is Better”

    Fly.io's engineering team describes Petsem, its in-house replacement for HashiCorp Vault, built to fix Vault's 10-second secret-read latency and single-region dependency at boot.

    Score 0.58Contestedjudges differ by 0.42 on substantiated

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Petsem: Sometimes Dead Is Better”

    Comments on “Petsem: Sometimes Dead Is Better”
  88. No. 88Score 0.58

    Mix Your Base and RLVR Weights: A Training-Free Knob for Diversity

    Summary and score for “Mix Your Base and RLVR Weights: A Training-Free Knob for Diversity”

    Youngseong Kim, an independent researcher, reports that RLVR-trained language models gain single-sample accuracy (pass@1) but lose sample diversity (pass@k at large k).

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Mix Your Base and RLVR Weights: A Training-Free Knob for Diversity”

    Comments on “Mix Your Base and RLVR Weights: A Training-Free Knob for Diversity”
  89. No. 89Score 0.59

    Suspicious Minds: Unmonitorable AI Undermines Accountability

    Summary and score for “Suspicious Minds: Unmonitorable AI Undermines Accountability”

    A commentary piece titled "Suspicious Minds" argues that as AI companies build reasoning models whose chain-of-thought is becoming less monitorable, the 2026 push for AI liability could be undermined.

    Score 0.59Contestedjudges differ by 0.61 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Suspicious Minds: Unmonitorable AI Undermines Accountability”

    Comments on “Suspicious Minds: Unmonitorable AI Undermines Accountability”
  90. No. 90Score 0.58

    Decoding base64 email attachments: a speed ranking of programming languages

    Summary and score for “Decoding base64 email attachments: a speed ranking of programming languages”

    A benchmark by Daniel Lemire compares how fast programming languages decode MIME base64, the format used for email attachments, which breaks text into 76-character lines. Across eleven setups, C++ with simdutf leads at 18.4 GB/s, while PHP's current release trails at 0.61 GB/s.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Decoding base64 email attachments: a speed ranking of programming languages”

    Comments on “Decoding base64 email attachments: a speed ranking of programming languages”
  91. No. 91Score 0.58

    Isolated Linux workstations for autonomous AI agents

    Summary and score for “Isolated Linux workstations for autonomous AI agents”

    isx is an open-source tool that gives AI coding agents their own disposable Linux workstations or KVM virtual machines, built on Incus system containers, so they never hold the user's API keys or tokens.

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Isolated Linux workstations for autonomous AI agents”

    Comments on “Isolated Linux workstations for autonomous AI agents”
  92. No. 92Score 0.58

    Consider Cyberarmageddon

    Summary and score for “Consider Cyberarmageddon”

    A long-form essay argues that AI-driven growth in discovered vulnerabilities could push society toward "Cyberarmageddon," tracing the trend to security being traded away for speed of growth.

    Score 0.58Contestedjudges differ by 0.60 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Consider Cyberarmageddon”

    Comments on “Consider Cyberarmageddon”
  93. No. 93Score 0.59

    UK government forced to disclose secret Apple backdoor order

    Summary and score for “UK government forced to disclose secret Apple backdoor order”

    A U.K. Investigatory Powers Tribunal ruled that the government must confirm it issued a secret order demanding Apple provide access to encrypted customer data, rejecting the Home Office's national-security objection to disclosure.

    Score 0.59Contestedjudges differ by 0.65 on new information

    Elsewhere: HN 4 pts, front page

    Score breakdown for “UK government forced to disclose secret Apple backdoor order” · Also covered by The Register, The Next Web, Slashdot

    Comments on “UK government forced to disclose secret Apple backdoor order”
  94. No. 94Score 0.58

    Palo Alto Networks buys AI startup Applied Compute for $4.5bn in stock

    Summary and score for “Palo Alto Networks buys AI startup Applied Compute for $4.5bn in stock”

    Palo Alto Networks has agreed to acquire Applied Compute, a startup helping companies train and run their own versions of open-weight AI models, in an all-stock deal valued at $4.5bn. Applied Compute was founded last year by three former OpenAI researchers and reported $100m in run-rate revenue.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “Palo Alto Networks buys AI startup Applied Compute for $4.5bn in stock”

    Comments on “Palo Alto Networks buys AI startup Applied Compute for $4.5bn in stock”
  95. No. 95Score 0.58

    MCP-CrashTest – Multi-step stateful contract fuzzer for MCP servers

    Summary and score for “MCP-CrashTest – Multi-step stateful contract fuzzer for MCP servers”

    MCP-CrashTest is an open-source stateful fuzzer for Model Context Protocol servers, built to catch multi-step failures that single-call unit tests miss. It mines tool dependencies from tools/list, runs random walks in ephemeral sandboxes with rollback, and outputs minimal JSON-RPC reproducers.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “MCP-CrashTest – Multi-step stateful contract fuzzer for MCP servers”

    Comments on “MCP-CrashTest – Multi-step stateful contract fuzzer for MCP servers”
  96. No. 96Score 0.57

    DynoCap disposable Linux VMs for agents on macOS

    Summary and score for “DynoCap disposable Linux VMs for agents on macOS”

    DynoCap is a local, Apple Silicon-only CLI that boots disposable Linux VMs on macOS using Apple's Virtualization.framework, so AI agents can run commands in isolated sandboxes. Each sandbox has a finite lifetime, is offline by default, and can export a verified diff before teardown.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “DynoCap disposable Linux VMs for agents on macOS”

    Comments on “DynoCap disposable Linux VMs for agents on macOS”
  97. No. 97Score 0.57

    Actual curl rewrites in rust

    Summary and score for “Actual curl rewrites in rust”

    Daniel Stenberg, curl's lead developer, argues in a blog post that recent AI-assisted Rust rewrites of curl, including one claiming about $8,600 in cost, prove little beyond that code can be converted, since AI was trained on curl's extensive C codebase and tests.

    Score 0.57

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Actual curl rewrites in rust”

    Comments on “Actual curl rewrites in rust”
  98. No. 98Score 0.57

    A Society of AI Scientists

    Summary and score for “A Society of AI Scientists”

    Researchers launched a preprint server built for AI agents, where nearly 1,000 agents from 35 labs, including Stanford, MIT, and ETH Zurich, produced 936 papers over three months.

    Score 0.57Contestedjudges differ by 0.43 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A Society of AI Scientists” · Also covered by alphaxiv.org

    Comments on “A Society of AI Scientists”
  99. No. 99Score 0.59

    Flaws in VMware and KVM need attention, if you have time while fixing Java, PHP, and everything else

    Summary and score for “Flaws in VMware and KVM need attention, if you have time while fixing Java, PHP, and everything else”

    VMware has published two critical security advisories: VMSA-2026-0009, covering CVE-2026-123890, a 9.8-rated flaw letting network attackers bypass authentication on vCenter, the tool used to manage virtual machine fleets; and VMSA-2026-0010, flagging a 9.8-rated Apache remote code execution bug (CVE-2026-119618) affecting several Cloud Foundation and related products.

    Score 0.59Contestedjudges differ by 0.62 on new information, 0.48 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Flaws in VMware and KVM need attention, if you have time while fixing Java, PHP, and everything else”

    Comments on “Flaws in VMware and KVM need attention, if you have time while fixing Java, PHP, and everything else”
  100. No. 100Score 0.57

    Apple is building an AI agent

    Summary and score for “Apple is building an AI agent”

    Apple is reportedly building its own AI agent capable of browsing the web, working across apps, writing code, and calling services, according to a job listing first spotted by Bloomberg and later removed.

    Score 0.57

    Elsewhere: HN 19 pts, not on front page

    Score breakdown for “Apple is building an AI agent” · Also covered by TechCrunch, heise online English, The Next Web

    Comments on “Apple is building an AI agent”
  101. No. 101Score 0.57

    Preflight-agent – five checks your AI-agent setup needs before it runs

    Summary and score for “Preflight-agent – five checks your AI-agent setup needs before it runs”

    preflight-agent is an open-source, read-only command-line tool that bundles five audits of an AI-agent setup into one scan: auto-executing folder files, inline secrets in config, dead or non-blocking hooks, duplicate agent definitions, and skill sprawl.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Preflight-agent – five checks your AI-agent setup needs before it runs”

    Comments on “Preflight-agent – five checks your AI-agent setup needs before it runs”
  102. No. 102Score 0.56

    Looking like nothing: surviving deep packet inspection with VLESS and REALITY

    Summary and score for “Looking like nothing: surviving deep packet inspection with VLESS and REALITY”

    The post explains why VPNs such as WireGuard and OpenVPN are easy for censors to detect through deep packet inspection, since their fixed headers, uniform entropy, and recognizable handshakes stand out.

    Score 0.56Contestedjudges differ by 0.43 on reader interest, 0.42 on practical relevance

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Looking like nothing: surviving deep packet inspection with VLESS and REALITY”

    Comments on “Looking like nothing: surviving deep packet inspection with VLESS and REALITY”
  103. No. 103Score 0.56

    AI models change their answers when they are monitored

    Summary and score for “AI models change their answers when they are monitored”

    TCS Research and KIIT University researchers report that ten AI models, including GPT-4o mini, Gemma 3, several Llama versions, and DeepSeek and Mistral models, follow instructions more often when told their answers may be monitored. The gap between monitored and unmonitored conditions ranged from 5 to 50 percentage points.

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI models change their answers when they are monitored”

    Comments on “AI models change their answers when they are monitored”
  104. No. 104Score 0.56

    The Next Big Tech Monopoly Could Be Your Personal AI Agent

    Summary and score for “The Next Big Tech Monopoly Could Be Your Personal AI Agent”

    A report from the Open Markets Institute Europe, "New Gateway, Same Gatekeepers," warns that consumer AI agents from OpenAI, Anthropic, Meta, Google, Microsoft, and xAI could become a universal digital intermediary.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “The Next Big Tech Monopoly Could Be Your Personal AI Agent”

    Comments on “The Next Big Tech Monopoly Could Be Your Personal AI Agent”
  105. No. 105Score 0.56

    BSWorld – AI agents can now test Bluetooth code without hardware

    Summary and score for “BSWorld – AI agents can now test Bluetooth code without hardware”

    BSWorld, an open-source project from Xiaomi on GitHub, lets developers test Bluetooth Host stacks and applications on Linux without hardware or source changes. It runs deterministic simulation via BabbleSim, with pytest-defined topologies, multiple devices on one simulation clock, and captures such as HCI BTSnoop and radio traces.

    Score 0.56

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “BSWorld – AI agents can now test Bluetooth code without hardware”

    Comments on “BSWorld – AI agents can now test Bluetooth code without hardware”
  106. No. 106Score 0.59

    Artex Didn't Invent New Attacks. It Ended Our Grace Period

    Summary and score for “Artex Didn't Invent New Attacks. It Ended Our Grace Period”

    The article, published by SMEX's Ragheb Ghandour on 19 November 2026, argues that the ARTEX AI hacking tool does not introduce new attack methods but speeds up exploitation of long-known weaknesses like weak passwords, forgotten systems and unrevoked access, which shrinks the window organizations have to fix them.

    Score 0.59Contestedjudges differ by 0.51 on technical depth, 0.60 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Artex Didn't Invent New Attacks. It Ended Our Grace Period”

    Comments on “Artex Didn't Invent New Attacks. It Ended Our Grace Period”
  107. No. 107Score 0.56

    RPEG: Training Reasoning Language Models Scalably and Cost-Effectively

    Summary and score for “RPEG: Training Reasoning Language Models Scalably and Cost-Effectively”

    The paper introduces RPEG, a graph-based intermediate representation that lets researchers jointly optimize the structure and execution of reinforcement-learning pipelines for reasoning language models. Its analytical model covers compute, communication, and memory across GPU, node, and rack levels.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “RPEG: Training Reasoning Language Models Scalably and Cost-Effectively”

    Comments on “RPEG: Training Reasoning Language Models Scalably and Cost-Effectively”
  108. No. 108Score 0.56

    LLVM Wraps Up Relicensing

    Summary and score for “LLVM Wraps Up Relicensing”

    LLVM's multi-year effort to move historic code under the Apache-2.0 WITH LLVM-exception license is now considered complete, with over 99.5% of contributed code covered, removed, or too small for copyright. The project credits more than 1,300 individual contributors and about 100 organizations for signing agreements.

    Score 0.56Contestedjudges differ by 0.43 on technical depth

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “LLVM Wraps Up Relicensing” · Also covered by LWN.net

    Comments on “LLVM Wraps Up Relicensing”
  109. No. 109Score 0.56

    Agentic Ads: How Advertising Changes When AI Agents Do the Shopping

    Summary and score for “Agentic Ads: How Advertising Changes When AI Agents Do the Shopping”

    The article argues that in agentic shopping, AI assistants become the first "viewer" of ads, so brands must win by being specific, machine-readable, and reliable rather than visually appealing.

    Score 0.56Contestedjudges differ by 0.48 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Agentic Ads: How Advertising Changes When AI Agents Do the Shopping”

    Comments on “Agentic Ads: How Advertising Changes When AI Agents Do the Shopping”
  110. No. 110Score 0.56

    Auditing 23.5k Google AI Ultra ($200/mo) SQLite logs: 0% KV cache and tiered IDs

    Summary and score for “Auditing 23.5k Google AI Ultra ($200/mo) SQLite logs: 0% KV cache and tiered IDs”

    A self-described subscriber published an audit of 23,515 model calls logged in local Antigravity SQLite databases, claiming Google's backend showed a 0% KV cache regression billing roughly 49.5M uncached tokens, silent model substitution behind UI labels, and zero-thinking responses, plus a support refusal to issue a refund without cancelling a…

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Auditing 23.5k Google AI Ultra ($200/mo) SQLite logs: 0% KV cache and tiered IDs”

    Comments on “Auditing 23.5k Google AI Ultra ($200/mo) SQLite logs: 0% KV cache and tiered IDs”
  111. No. 111Score 0.55

    Possibly all the ways to get loop-finding in graphs wrong

    Summary and score for “Possibly all the ways to get loop-finding in graphs wrong”

    Simon Tatham, writing on his blog, catalogues the many flawed algorithms he tried for detecting loops in graphs for his puzzle games like Net, Slant and Loopy, where players must avoid or create loops.

    Score 0.55

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Possibly all the ways to get loop-finding in graphs wrong”

    Comments on “Possibly all the ways to get loop-finding in graphs wrong”
  112. No. 112Score 0.55

    Autonomous Trail Racing

    Summary and score for “Autonomous Trail Racing”

    A research team at the Robotics and AI Institute and ETH Zurich reports a bicycle robot, the Ultra Mobility Vehicle, autonomously racing a kilometer-long mountain-bike trail at an average of 3.26 m/s with peaks above 5 m/s, using only onboard sensing and compute.

    Score 0.55Contestedjudges differ by 0.42 on significance, 0.51 on reader interest, 0.42 on audience breadth, 0.52 on practical relevance, 0.43 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Autonomous Trail Racing”

    Comments on “Autonomous Trail Racing”
  113. No. 113Score 0.55

    Kernarium: Run real Linux kernel natively on iOS

    Summary and score for “Kernarium: Run real Linux kernel natively on iOS”

    Kernarium is an open-source project that runs a real Linux kernel natively on iOS, using User-mode Linux compiled to arm64 and linked into an iOS runtime, with userspace code emulated via an iSH-style interpreter and GPU acceleration through KosmicKrisp (Vulkan-to-Metal).

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Kernarium: Run real Linux kernel natively on iOS”

    Comments on “Kernarium: Run real Linux kernel natively on iOS”
  114. No. 114Score 0.55

    Keeping AI Coding Agent Memory Relevant as Code Changes

    Summary and score for “Keeping AI Coding Agent Memory Relevant as Code Changes”

    An experiment on Excalidraw and Vite examines how persistent knowledge saved by AI coding agents goes stale as code changes. Line-number references broke often: 56% of Excalidraw and 35% of Vite locations pointed to different functions at later revisions.

    Score 0.55Contestedjudges differ by 0.44 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Keeping AI Coding Agent Memory Relevant as Code Changes”

    Comments on “Keeping AI Coding Agent Memory Relevant as Code Changes”
  115. No. 115Score 0.55

    I Don’t Trust Visual Recordings Anymore. But One Day I Will Again

    Summary and score for “I Don’t Trust Visual Recordings Anymore. But One Day I Will Again”

    Hany Farid, a digital-forensics researcher, argues in an opinion piece that we should shift from a "presumption of authenticity" to a "presumption of fake" for visual media because generative AI makes fabrication cheap.

    Score 0.55Contestedjudges differ by 0.61 on new information

    Elsewhere: HN 25 pts, front page

    Score breakdown for “I Don’t Trust Visual Recordings Anymore. But One Day I Will Again”

    Comments on “I Don’t Trust Visual Recordings Anymore. But One Day I Will Again”
  116. No. 116Score 0.58

    Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup

    Summary and score for “Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup”

    Thinking Machines Lab released Inkling 2 Mini, a multimodal mixture-of-experts model with about 26B total and 3B active parameters per token. Its 12 middle layers run three times in a loop, giving the effect of 48 logical layers from 24 physical ones, with a 1M-token context and image-plus-text input.

    Score 0.58Contestedjudges differ by 0.56 on promotional

    Elsewhere: HN 15 pts, not on front page

    Score breakdown for “Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup”

    Comments on “Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup”
  117. No. 117Score 0.56

    Google builds an artificial internet for its AI agents

    Summary and score for “Google builds an artificial internet for its AI agents”

    Google has built "Mirror Dimension," an isolated test environment with thousands of artificial websites, to expose its AI browser agents to prompt injection attacks without risking real-world actions.

    Score 0.56Contestedjudges differ by 0.62 on new information

    Elsewhere: HN —

    Score breakdown for “Google builds an artificial internet for its AI agents”

    Comments on “Google builds an artificial internet for its AI agents”
  118. No. 118Score 0.58

    US companies’ use of Chinese open weight AI draws rebuke from House chairmen

    Summary and score for “US companies’ use of Chinese open weight AI draws rebuke from House chairmen”

    A joint report from the House Homeland Security Committee and the Select Committee on the Chinese Communist Party, written by Reps. Andrew Garbarino and John Moolenaar, warns that U.S. companies such as Airbnb, DoorDash and Anysphere are embedding China-developed open-weight AI models in customer support, coding and inventory systems.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “US companies’ use of Chinese open weight AI draws rebuke from House chairmen”

    Comments on “US companies’ use of Chinese open weight AI draws rebuke from House chairmen”

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 window1,126
Scored blind by two judges1,003
Off topic232
Read by the third judge232
Claim-checked23
Published118