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 129 stories · 11 hidden by your filters

  1. No. 1Score 0.80

    PHP-FPM request smuggling allows unauthenticated remote code execution

    Summary and score for “PHP-FPM request smuggling allows unauthenticated remote code execution”

    A PHP security advisory describes a PHP-FPM request smuggling flaw, fixed in the 19 November releases (8.5.13, 8.4.28, 8.3.37, 8.2.36). Unread php://input bytes can be parsed as a forged FastCGI request, letting an unauthenticated attacker run PHP as the worker. It requires persistent backend connections, such as nginx fastcgi_keep_conn, and an app that opens php://input without reading it fully. The affected-since-5.6.0 claim rests on code inspection. Until upgrading, disable connection reuse. LSAPI is affected by desynchronisation only.

    Score 0.80Claim check: 8 no source given

    Elsewhere: HN 5 pts, not on front page

    Score breakdown and claims for “PHP-FPM request smuggling allows unauthenticated remote code execution” · Also covered by aikido.dev, php.net, depthfirst, GitHub, GitHub, heise online English

    Comments on “PHP-FPM request smuggling allows unauthenticated remote code execution”
  2. No. 2Score 0.79

    Ruby2Shell (GHSA-34m4-m9fh-gqp5)

    Summary and score for “Ruby2Shell (GHSA-34m4-m9fh-gqp5)”

    Veria Labs disclosed GHSA-34m4-m9fh-gqp5, a one-byte heap overflow in Ruby's case mapping for single-byte encodings such as ISO-8859 and Windows-125x. Upcasing ß writes two bytes but the loop checks for one, so 21 of them overrun a buffer sized at input length plus 20. Rails and classic Sinatra are reachable by default. Fixed in Ruby 3.3.14, 3.4.13 and 4.0.10. Per Veria, GitLab, GitHub Enterprise Server and Chatwoot have no fix yet; a Rails remote code execution proof of concept is due 25 November. Most claims here come from Veria alone.

    Score 0.79Claim check: 1 matches source · 7 no source given

    Elsewhere: HN 391 pts, front page

    Score breakdown and claims for “Ruby2Shell (GHSA-34m4-m9fh-gqp5)” · Also covered by Durendal, heise online English

    Comments on “Ruby2Shell (GHSA-34m4-m9fh-gqp5)”
  3. No. 3Score 0.78

    De-fanging dangerous triggers in GitHub Actions across an organization

    Summary and score for “De-fanging dangerous triggers in GitHub Actions across an organization”

    Alpha-Omega and the Python Software Foundation published a guide for locking down risky GitHub Actions triggers, ahead of GitHub blocking pull_request_target by default in public repositories on December 2, 2026. The authors, PyPI security engineer Mike Fiedler and Seth Larson, recommend an organization policy that allowlists seven events, first run in Evaluate mode, and they list likely breakage such as labelers and slash-command bots. They cite incidents including Ultralytics, LiteLLM, Telnyx and elementary-data. Free plans need per-repository policies, and registry checks alone miss tag-push routes.

    Score 0.78Claim check: 7 matches source · 1 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “De-fanging dangerous triggers in GitHub Actions across an organization” · Also covered by Python Software Foundation Blog

    Comments on “De-fanging dangerous triggers in GitHub Actions across an organization”
  4. No. 4Score 0.77

    PostgreSQL 19

    Summary and score for “PostgreSQL 19”

    PostgreSQL 19's release notes list REPACK, including a CONCURRENTLY mode that avoids blocking reads and writes, plus the pg_plan_advice extension for steering the planner, parallel autovacuum, and a WAIT command for read-your-writes on standbys. Several defaults change: JIT is off, lz4 replaces pglz for TOAST, and max_locks_per_transaction doubles to 128. RADIUS authentication is removed, and standard_conforming_strings is always on. Upgrades need pg_upgrade, dump/restore or logical replication. Some notes, including the release date and the wraparound-warning and C11 claims, could not be verified.

    Score 0.77Claim check: 1 matches source · 1 no source given · 6 couldn’t check

    Elsewhere: HN 13 pts, not on front page

    Score breakdown and claims for “PostgreSQL 19”

    Comments on “PostgreSQL 19”
  5. No. 5Score 0.77

    Hardening GitHub Actions Against Insider Threats

    Summary and score for “Hardening GitHub Actions Against Insider Threats”

    Socket's Karlo Zanki publishes the second half of a GitHub Actions hardening guide, covering attackers who hold a valid credential, such as a phished maintainer or a departing insider. Ten chapters cover security keys, short-lived tokens, push rulesets on workflow files, execution protections, npm trusted publishing, staged releases and a revocation drill. It frames them as containment and detection, not prevention. Its LiteLLM account describes one cascading compromise, so the number to raise is independently protected credentials, not identities. The five-second override detail could not be verified.

    Score 0.77Contestedjudges differ by 0.48 on new informationClaim check: 1 matches source · 1 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Hardening GitHub Actions Against Insider Threats”

    Comments on “Hardening GitHub Actions Against Insider Threats”
  6. No. 6Score 0.73

    A Megakernel per Model

    Summary and score for “A Megakernel per Model”

    Xingyu Dang reports that Claude coding agents wrote a single-GPU-program "megakernel" for each of eight open models, from 8B to 594B parameters. At batch size 1, with the reference data flow, the engines run 1.13–1.77× faster than the fastest of vLLM, SGLang or Cohere's own megakernel; allowing lossless storage changes raises this to 2.05×. On a production trace with Inkling-2-Small on 4×B200, 99–100% of requests met their latency targets versus 82–85% for SGLang. Gains shrink at batch size 8, and the post's own figures cover specific models and hardware.

    Score 0.73Claim check: 3 matches source · 4 no source given · 1 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “A Megakernel per Model”

    Comments on “A Megakernel per Model”
  7. No. 7Score 0.73

    PHP 8.6 Release Announcement

    Summary and score for “PHP 8.6 Release Announcement”

    PHP 8.6, released November 19, 2026 according to php.net, adds partial function application with ? placeholders, a \Time\Duration class, an Io\Poll API built on epoll, kqueue and similar mechanisms for event loops like ReactPHP and AMPHP, clamp(), a SortDirection enum, and readonly property defaults. Upgrade work centres on the deprecations: mbstring regex functions, return inside finally, and several aliases. Sessions also become stricter by default, with strict mode and HttpOnly on and SameSite set to Lax. The announcement gives no independent benchmarks.

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

    Elsewhere: HN 284 pts, front page

    Score breakdown and claims for “PHP 8.6 Release Announcement”

    Comments on “PHP 8.6 Release Announcement”
  8. No. 8Score 0.75

    Reaching Native Database Attack Surface from Minimal SQL Roles

    Summary and score for “Reaching Native Database Attack Surface from Minimal SQL Roles”

    Security firm depthfirst reports remote-code-execution chains from minimally privileged SQL accounts in PostgreSQL and MySQL, plus a reproduced ClickHouse overflow and an Oracle AI Database flaw (CVE-2026-114819, CVSS 9.9). The PostgreSQL chain relies on default temporary-schema rights; MySQL and ClickHouse need only CREATE TEMPORARY TABLES. The point for builders is that narrow grants do not shrink the reachable C/C++ engine code. Most specifics, including the CVEs and ClickHouse's default setting, were not independently verified.

    Score 0.75Claim check: 1 matches source · 2 no source given · 5 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Reaching Native Database Attack Surface from Minimal SQL Roles”

    Comments on “Reaching Native Database Attack Surface from Minimal SQL Roles”
  9. No. 9Score 0.72

    PostgreSQL 19 Released

    Summary and score for “PostgreSQL 19 Released”

    PostgreSQL 19 shipped on November 19, about two months later than usual, which Phoronix and Simon Willison attribute to the workload of handling AI-reported bugs and vulnerabilities. Headline changes are REPACK (CONCURRENTLY), which reclaims table bloat without blocking reads or writes, the pg_plan_advice extension for fixing query plans, and WAIT FOR LSN for read-your-writes on standbys. JIT is now off by default and RADIUS authentication is removed. The project's "nearly 2x" bulk-insert gain from faster foreign key checks comes from its own testing, with no details given.

    Score 0.72Claim check: 3 matches source · 3 no source given · 2 couldn’t check

    Elsewhere: HN 101 pts, front page

    Score breakdown and claims for “PostgreSQL 19 Released” · Also covered by Phoronix, LWN.net, Simon Willison

    Comments on “PostgreSQL 19 Released”
  10. No. 10Score 0.69

    Fable 5.2 vs. PCB routing and placing

    Summary and score for “Fable 5.2 vs. PCB routing and placing”

    PCBJam gave Fable 5.2 stripped KiCad boards (copper removed, originals hidden) and compared the results with human layouts, scored by Claude Opus 5.5 and by its own scripts. Small boards came out DRC-clean, and the Glasgow revC3 routing run connected every net. The ESP32-PoE and OrangeCrab runs ended with open links, and the M.2 adapter run had zero pair skew. A purpose-built CLI helped more than an MCP server. The figures are PCBJam's own and unverified, and reviewer scores come from an AI model.

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

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Fable 5.2 vs. PCB routing and placing”

    Comments on “Fable 5.2 vs. PCB routing and placing”
  11. No. 11Score 0.69

    Five Eligible CPUs, a Four-CPU Quota: What Runtimes Report and What It Costs

    Summary and score for “Five Eligible CPUs, a Four-CPU Quota: What Runtimes Report and What It Costs”

    Octri's Emad Khair compares what runtimes report in a container with five eligible CPUs but a four-CPU quota: Python 3.13 and Go 1.23 say five, Node.js 22 and OpenJDK 21 say four. In a small Python test, five CPU-bound workers instead of four left batch time near 1.0 s but raised 95th-percentile job time from 50.4 to 70.9 ms, with throttling in nine of ten periods. The workload is narrow, but it supports setting worker counts explicitly. Go 1.25 now considers cgroup quotas in its default GOMAXPROCS.

    Score 0.69Claim check: 4 matches source · 4 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Five Eligible CPUs, a Four-CPU Quota: What Runtimes Report and What It Costs”

    Comments on “Five Eligible CPUs, a Four-CPU Quota: What Runtimes Report and What It Costs”
  12. No. 12Score 0.68

    Recursive Sparse Attention: Million-Token Extrapolation at Logarithmic Cost

    Summary and score for “Recursive Sparse Attention: Million-Token Extrapolation at Logarithmic Cost”

    Haoai Lab's RecSA builds a tree over the KV cache so each query searches it level by level, keeping selection cost and entries read at O(log L). Unselected nodes stay in the softmax as summaries, and candidate-local RoPE keeps relative distances within the trained range. Starting from Qwen3-4B and training on 0.62B tokens at 32K, it scores 60.5 on RULER at 1M versus 40.0 for Dense + DCA, and decodes 3.8× faster at 524K on one H200. Results are the authors' own, on a 4B model; larger models are untested.

    Score 0.68Claim check: 2 matches source · 6 no source given

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “Recursive Sparse Attention: Million-Token Extrapolation at Logarithmic Cost”

    Comments on “Recursive Sparse Attention: Million-Token Extrapolation at Logarithmic Cost”
  13. No. 13Score 0.67

    Design for Understandability

    Summary and score for “Design for Understandability”

    Jonathan Schuster, in a talk given at Google's internal Code Health Summit in November 2025, argues that as LLMs write more code, engineers' main job shifts to reading and judging it, so systems should be designed for understandability, as they were once designed for testability. He proposes three practices: design by contract, data invariants, and isolating side effects and global state, all aimed at local reasoning. The claims are argued from examples and experience, not measured, and several lack cited sources.

    Score 0.67Contestedjudges differ by 0.40 on substantiatedClaim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Design for Understandability”

    Comments on “Design for Understandability”
  14. No. 14Score 0.67

    valkey-io/valkey 9.2.0

    Summary and score for “valkey-io/valkey 9.2.0”

    Valkey 9.2.0, the first stable release of the 9.2 line, is rated low upgrade urgency in its release notes. Headline changes include opt-in forkless RDB snapshots and forkless full sync to replicas, Zstd streaming compression for RDB and replication, and a B+ tree replacing the skiplist for large sorted sets (OBJECT ENCODING now reports btree). It also adds the Path Hash type, ACL roles, hot key detection and the DELEX command. Operators should note that active expiration now marks data dirty, so BGSAVE may trigger more often.

    Score 0.67Claim check: 7 matches source · 1 no source given

    Elsewhere: HN —

    Score breakdown and claims for “valkey-io/valkey 9.2.0”

    Comments on “valkey-io/valkey 9.2.0”
  15. No. 15Score 0.67

    Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web

    Summary and score for “Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web”

    Google's Chrome security team describes a "mirror dimension," a synthetic web for testing its auto browse agent. Headless Chrome runs in a gVisor sandbox, and a proxy mints TLS certificates so agents reach real URLs that resolve to generated sites, which avoids tipping them off. Payloads for prompt-injection tests travel in HTTP headers, not URL parameters. Google reports over 1,000 test candidates per engineer daily, but these figures and the claim that such sandboxes will become the baseline are the company's own, with no independent sourcing.

    Score 0.67Claim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web”

    Comments on “Mirror Dimension for Agents: Evaluating Browser-Use Agents Without the Web”
  16. No. 16Score 0.66

    Distributed, high-performance blob storage on SQLite

    Summary and score for “Distributed, high-performance blob storage on SQLite”

    Zerodha's engineers describe Blobasaur, an open-source Rust server that speaks the Redis protocol and stores JSON blobs in sharded SQLite files, replacing the unmaintained SSDB behind its Kite platform. In the authors' benchmark on a single laptop with tmpfs, 16 shards raised write throughput about 3.3x over one shard (25,324 vs 7,740 ops/s), while reads were unchanged. In production, a three-node cluster reports p99 near 2 ms. Durability is relaxed, and slot rebalancing and failover are not yet implemented. Most figures are self-reported.

    Score 0.66Claim check: 1 matches source · 5 no source given · 2 couldn’t check

    Elsewhere: HN 277 pts, front page

    Score breakdown and claims for “Distributed, high-performance blob storage on SQLite”

    Comments on “Distributed, high-performance blob storage on SQLite”
  17. No. 17Score 0.73

    Three breaches, endless leaks: Your AI supply chain needs canaries

    Summary and score for “Three breaches, endless leaks: Your AI supply chain needs canaries”

    Dragos researcher HD Moore says about two million canary-seeded prompts, sent to AI services and routers from July to November 2026, exposed theft at three hosting providers, where bait AWS keys drew tens of thousands of API calls. He also reports marketplace resellers logging prompts and a Chinese lab fetching URLs despite Zero Data Retention flags. The findings are Dragos's own and unverified externally, but they suggest privacy flags bind only the first hop, so teams should seed their own canaries.

    Score 0.73Contestedjudges differ by 0.46 on promotionalClaim check: 6 no source given · 2 couldn’t check

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “Three breaches, endless leaks: Your AI supply chain needs canaries”

    Comments on “Three breaches, endless leaks: Your AI supply chain needs canaries”
  18. No. 18Score 0.67

    Company docs didn't help our AI agent. Four shortnotes doubled accuracy

    Summary and score for “Company docs didn't help our AI agent. Four shortnotes doubled accuracy”

    Contextflo ran a Claude Haiku 5.5 agent on DABstep, the Adyen and Hugging Face payments benchmark, and found that company docs did not raise accuracy while four short rule notes did. On 42 held-out questions, tries right went from 40% (tables only) to 36% (docs) to 76% (docs plus notes), at roughly a quarter of the cost. One broad note broke a previously correct answer type. The results are self-reported, use one model and one dataset, and mix official and rebuilt scoring.

    Score 0.67Claim check: 1 matches source · 7 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Company docs didn't help our AI agent. Four shortnotes doubled accuracy”

    Comments on “Company docs didn't help our AI agent. Four shortnotes doubled accuracy”
  19. No. 19Score 0.64

    UK court confirms British government served Apple with secret encryption notice

    Summary and score for “UK court confirms British government served Apple with secret encryption notice”

    The UK's Investigatory Powers Tribunal has ruled that the government served Apple with a technical capability notice, rejecting its "neither confirm nor deny" policy. The judges said earlier disclosures to reporters and US officials had already revealed the notice, and MacRumors reports they called the national security argument "fanciful." The notice reportedly concerns iCloud's Advanced Data Protection, which Apple withdrew for UK users last year. The ruling does not decide lawfulness; a hearing is set for early 2027. It shows how secret orders can pressure end-to-end encryption design.

    Score 0.64Claim check: 4 matches source · 3 no source given · 1 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “UK court confirms British government served Apple with secret encryption notice” · Also covered by MacAnorak, Privacy International, MacRumors, heise online English

    Comments on “UK court confirms British government served Apple with secret encryption notice”
  20. No. 20Score 0.66

    The fix can't wait. What breaks when it lands?

    Summary and score for “The fix can't wait. What breaks when it lands?”

    IR Labs, which is building fix-verification tooling, traced the Linux kernel's fix graph and estimates that about one runtime fix in thirty (3.40%) broke something that needed its own repair within a year. The figure comes from a model-assisted audit of 200 sampled cases and has no independent source. From a 2024 mmap patch alone, its tool wrote tests that flagged a sealed-file regression. In early results it found 4 of 6 real breaks and 7 of 10 planted defects, small samples.

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

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “The fix can't wait. What breaks when it lands?”

    Comments on “The fix can't wait. What breaks when it lands?”
  21. No. 21Score 0.64

    Velarx: A container engine with a chain of trust from signature to page reads

    Summary and score for “Velarx: A container engine with a chain of trust from signature to page reads”

    Velarx is a new container engine written in Rust, built around composefs and designed to verify images from their signature down to each page the kernel reads.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Velarx: A container engine with a chain of trust from signature to page reads”

    Comments on “Velarx: A container engine with a chain of trust from signature to page reads”
  22. No. 22Score 0.64

    A Linux time travel system

    Summary and score for “A Linux time travel system”

    clifr is an open-source (AGPL-3.0) Linux "flight recorder" from developer jlruedab, currently a 0.1.0 pre-release. A background service, clifrd, continuously records processes, network connections, files, mounts and kernel messages using eBPF, keeping 72 hours by default.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A Linux time travel system”

    Comments on “A Linux time travel system”
  23. No. 23Score 0.76

    Microsoft's Agent Framework PyPI package compromised to steal credentials

    Summary and score for “Microsoft's Agent Framework PyPI package compromised to steal credentials”

    Attackers compromised Microsoft's official Agent Framework package on PyPI, publishing versions 1.26.1 and 1.26.2 that ran a credential stealer at Python startup, even without importing the package. The malware targeted cloud and CI credentials and sent them to an attacker-controlled endpoint.

    Score 0.76Contestedjudges differ by 0.59 on new informationClaim 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 “Microsoft's Agent Framework PyPI package compromised to steal credentials”

    Comments on “Microsoft's Agent Framework PyPI package compromised to steal credentials”
  24. No. 24Score 0.66

    ChatGPT Helped "Sadistic" School Shooter Build 3D-Printed Weapons

    Summary and score for “ChatGPT Helped "Sadistic" School Shooter Build 3D-Printed Weapons”

    Mother Jones reporter Mark Follman reviewed Jesse Van Rootselaar's ChatGPT history and found that the 2026 Tumbler Ridge school shooter used the chatbot repeatedly between June 2025 and February 2026 for firearm and magazine advice, homemade explosives, suicide methods, and violent fantasies.

    Score 0.66Claim check: 2 no source given · 6 couldn’t check

    Elsewhere: HN 6 pts, not on front page

    Score breakdown and claims for “ChatGPT Helped "Sadistic" School Shooter Build 3D-Printed Weapons”

    Comments on “ChatGPT Helped "Sadistic" School Shooter Build 3D-Printed Weapons”
  25. No. 25Score 0.70

    Vite Remote Code Execution

    Summary and score for “Vite Remote Code Execution”

    A security advisory for Vite, the JavaScript build tool, describes a remote code execution flaw in its dev server. Crafted HTTP requests could inject code into virtual modules, allowing reads of files outside the allow list and arbitrary server-side evaluation.

    Score 0.70Claim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Vite Remote Code Execution”

    Comments on “Vite Remote Code Execution”
  26. No. 26Score 0.72

    AI-found Java flaw lets attackers into Jira and Jenkins without a login

    Summary and score for “AI-found Java flaw lets attackers into Jira and Jenkins without a login”

    Security start-up Calif disclosed Java4Shell, a critical Java flaw that can let attackers run code on Jira, Jenkins and other widely used business software without logging in.

    Score 0.72

    Elsewhere: HN —

    Score breakdown for “AI-found Java flaw lets attackers into Jira and Jenkins without a login” · Also covered by Engadget

    Comments on “AI-found Java flaw lets attackers into Jira and Jenkins without a login”
  27. No. 27Score 0.70

    AWS CVEs 2024-2026: Is AI Finding the Bugs, or Writing Them?

    Summary and score for “AWS CVEs 2024-2026: Is AI Finding the Bugs, or Writing Them?”

    A researcher traced AWS security bulletins from 2024 through November 2026, finding CVEs rose from 15 to 184 and that outside reporters, including AI firms like Anthropic, increasingly get credited.

    Score 0.70

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AWS CVEs 2024-2026: Is AI Finding the Bugs, or Writing Them?”

    Comments on “AWS CVEs 2024-2026: Is AI Finding the Bugs, or Writing Them?”
  28. No. 28Score 0.65

    Building iOS apps on Windows: what it takes

    Summary and score for “Building iOS apps on Windows: what it takes”

    OpenFruit describes how it builds unmodified Xcode projects into signed iOS apps natively on Windows, without patching sources or project files. Its engineering notes cover the hard parts: resolving Xcode build settings and conditional layers, extracting Apple's iOS SDK from a downloaded .xip, running Apple's original Swift macro binaries on…

    Score 0.65

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Building iOS apps on Windows: what it takes”

    Comments on “Building iOS apps on Windows: what it takes”
  29. No. 29Score 0.63

    I ran one Nginx log with 43,231 visitors through five log analyzers

    Summary and score for “I ran one Nginx log with 43,231 visitors through five log analyzers”

    A controlled test of five open-source nginx log analyzers on a synthetic log with 43,231 known people found that unique visitor counts ranged from 49,616 to 105,885, and still differed (49,576 to 60,209) after bots were removed.

    Score 0.63Claim check: 7 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “I ran one Nginx log with 43,231 visitors through five log analyzers”

    Comments on “I ran one Nginx log with 43,231 visitors through five log analyzers”
  30. No. 30Score 0.69

    Right Code, Wrong Comment: The Pointer Chase Inside Zlib

    Summary and score for “Right Code, Wrong Comment: The Pointer Chase Inside Zlib”

    A detailed write-up examines zlib's deflate internals and a Lean 4 formal safety proof of zlib 1.4. It explains that the two-pass design lets the output buffer pending_buf overlap the unread symbol buffer sym_buf, and that the release-build assert guarding this overlap compiles away.

    Score 0.69

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Right Code, Wrong Comment: The Pointer Chase Inside Zlib”

    Comments on “Right Code, Wrong Comment: The Pointer Chase Inside Zlib”
  31. No. 31Score 0.68

    VMSA-2026-0009: VMware vCenter, Tools, and SDDC Manager updates vulnerabilities

    Summary and score for “VMSA-2026-0009: VMware vCenter, Tools, and SDDC Manager updates vulnerabilities”

    Broadcom's VMSA-2026-0009 advisory, dated 19 November 2026, covers five vulnerabilities across VMware vCenter, SDDC Manager, Tools, and related Cloud Foundation and Telco products. The headline issue is CVE-2026-123890, a critical authentication bypass in the VMware Directory Service rated 9.8 CVSS, with no workaround available.

    Score 0.68

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “VMSA-2026-0009: VMware vCenter, Tools, and SDDC Manager updates vulnerabilities” · Also covered by Google Threat Intelligence (Mandiant)

    Comments on “VMSA-2026-0009: VMware vCenter, Tools, and SDDC Manager updates vulnerabilities”
  32. No. 32Score 0.65

    Inference After Coding Agents

    Summary and score for “Inference After Coding Agents”

    This paper argues that AI coding agents, which can search open-endedly over analysis choices and then have their code replayed, create a new form of post-selection inference that breaks classical p-value guarantees.

    Score 0.65Contestedjudges differ by 0.46 on reader interest

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Inference After Coding Agents”

    Comments on “Inference After Coding Agents”
  33. No. 33Score 0.59

    Microsoft will include PowerShell 7 in future Windows releases

    Summary and score for “Microsoft will include PowerShell 7 in future Windows releases”

    Microsoft plans to ship PowerShell 7 with future Windows releases and has announced that Windows PowerShell 5.1 is deprecated, though it stays supported and installed for now.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “Microsoft will include PowerShell 7 in future Windows releases”

    Comments on “Microsoft will include PowerShell 7 in future Windows releases”
  34. No. 34Score 0.65

    Detecting Empty Abstractions and Pruning Dead Paths in the Linux eBPF Verifier

    Summary and score for “Detecting Empty Abstractions and Pruning Dead Paths in the Linux eBPF Verifier”

    A paper proposes a constant-time emptiness check for the Linux eBPF verifier's abstract domains (intervals plus tristate-number bitmasks) to detect and prune unreachable branches. Refining a register's state under an infeasible condition can yield contradictory domains whose intersection is empty, violating invariants that downstream operators assume and causing production outages…

    Score 0.65

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Detecting Empty Abstractions and Pruning Dead Paths in the Linux eBPF Verifier”

    Comments on “Detecting Empty Abstractions and Pruning Dead Paths in the Linux eBPF Verifier”
  35. No. 35Score 0.59

    Intel signs first external customer for advanced 18A chips

    Summary and score for “Intel signs first external customer for advanced 18A chips”

    Intel CEO Lip-Bu Tan said at the TomorrowX Summit that the company has signed its first external customer for the 18A process, a milestone in Intel's effort to become a competitive contract chipmaker against TSMC.

    Score 0.59Contestedjudges differ by 0.47 on reader interest

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Intel signs first external customer for advanced 18A chips” · Also covered by The Next Web

    Comments on “Intel signs first external customer for advanced 18A chips”
  36. No. 36Score 0.58

    IOSCO: financial regulators must prepare for quantum impact

    Summary and score for “IOSCO: financial regulators must prepare for quantum impact”

    IOSCO, the global forum of securities regulators, published a final report warning that a sufficiently capable quantum computer could break public-key cryptography underpinning financial markets.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “IOSCO: financial regulators must prepare for quantum impact”

    Comments on “IOSCO: financial regulators must prepare for quantum impact”
  37. No. 37Score 0.66

    On The Use of AI-Based Tools to Find Bugs in Open Source Software

    Summary and score for “On The Use of AI-Based Tools to Find Bugs in Open Source Software”

    HAProxy lead developer Willy Tarreau describes his experience using AI tools for security bug-hunting in open source software. He reports that AI-generated reports often lack reproducers or patches, and that restricted access to Anthropic's Mythos model and frequent safeguard refusals in other Claude models hinder defensive security work.

    Score 0.66

    Elsewhere: HN 9 pts, not on front page

    Score breakdown for “On The Use of AI-Based Tools to Find Bugs in Open Source Software”

    Comments on “On The Use of AI-Based Tools to Find Bugs in Open Source Software”
  38. No. 38Score 0.61

    Text, Filters, and Now Vectors in One Postgres Index

    Summary and score for “Text, Filters, and Now Vectors in One Postgres Index”

    ParadeDB announced native vector support inside the same index that powers its Postgres text search, so filters and vectors are evaluated together rather than in separate passes.

    Score 0.61Contestedjudges differ by 0.45 on promotional

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Text, Filters, and Now Vectors in One Postgres Index”

    Comments on “Text, Filters, and Now Vectors in One Postgres Index”
  39. No. 40Score 0.65

    BSI raises threat level to “tense to critical”

    Summary and score for “BSI raises threat level to “tense to critical””

    Germany's Federal Office for Information Security (BSI) has raised its threat assessment to "tense to critical," its first such escalation since 2023, citing a surge in newly disclosed vulnerabilities and a shrinking window of about eight days, or 192 hours, between disclosure and exploitation in the second quarter of 2026.…

    Score 0.65

    Elsewhere: HN —

    Score breakdown for “BSI raises threat level to “tense to critical””

    Comments on “BSI raises threat level to “tense to critical””
  40. No. 41Score 0.65

    AI Agent Authorization That Limits the Damage of Prompt Injection

    Summary and score for “AI Agent Authorization That Limits the Damage of Prompt Injection”

    A Quito Lambda session demonstration by Jonathan Puglla, a Stack Builders engineer, showed how an AI agent can be induced by prompt injection in a ticket to close unrelated incidents when it holds broad credentials.

    Score 0.65

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI Agent Authorization That Limits the Damage of Prompt Injection”

    Comments on “AI Agent Authorization That Limits the Damage of Prompt Injection”
  41. No. 42Score 0.65

    Lightweight Fault Isolation (LFI) in LLVM

    Summary and score for “Lightweight Fault Isolation (LFI) in LLVM”

    LFI (Lightweight Fault Isolation) is an LLVM-based sandboxing technique that compiles existing C/C++ and assembly code to run in-process within a 4GiB sandbox region, with no source changes needed.

    Score 0.65

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Lightweight Fault Isolation (LFI) in LLVM”

    Comments on “Lightweight Fault Isolation (LFI) in LLVM”
  42. No. 43Score 0.60

    FailTrace – Reproduce a failed GitHub Actions run locally from its URL

    Summary and score for “FailTrace – Reproduce a failed GitHub Actions run locally from its URL”

    FailTrace is a command-line tool, now in early development, that takes the URL of a failed GitHub Actions run and reports which job and step failed, matched to the workflow file at that exact commit with the relevant error lines pulled from the logs.

    Score 0.60

    Elsewhere: HN 31 pts, front page

    Score breakdown for “FailTrace – Reproduce a failed GitHub Actions run locally from its URL”

    Comments on “FailTrace – Reproduce a failed GitHub Actions run locally from its URL”
  43. No. 44Score 0.64

    Autonomous AI Security Auditing of Webmin: 507 Newly Discovered Vulnerabilities

    Summary and score for “Autonomous AI Security Auditing of Webmin: 507 Newly Discovered Vulnerabilities”

    An autonomous Codex campaign, designed by an experienced security researcher, audited Webmin 2.670 at a pinned source revision and produced 507 reproduced, canonical vulnerability findings, each with CVSS 4.0 Base assessments and explicit applicability conditions. The author estimates the audit captured over 90% of classic defects at that revision.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Autonomous AI Security Auditing of Webmin: 507 Newly Discovered Vulnerabilities”

    Comments on “Autonomous AI Security Auditing of Webmin: 507 Newly Discovered Vulnerabilities”
  44. No. 45Score 0.56

    The Sanders-Casar AI Ban Would Turn a Forecast into a Felony

    Summary and score for “The Sanders-Casar AI Ban Would Turn a Forecast into a Felony”

    Reason's Jack Nicastro-style critique argues that the Sanders-Casar Ban Artificial Superintelligence Act would criminalize AI systems based on what they could do rather than harm they have caused.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The Sanders-Casar AI Ban Would Turn a Forecast into a Felony”

    Comments on “The Sanders-Casar AI Ban Would Turn a Forecast into a Felony”
  45. No. 46Score 0.64

    I rebuilt evilgrade to test where a leaked CA still works

    Summary and score for “I rebuilt evilgrade to test where a leaked CA still works”

    A research write-up describes rebuilding the evilgrade tool to test how far a leaked eMudhra intermediate CA, cross-chaining to the AAA root, can be exploited via TLS-only man-in-the-middle attacks.

    Score 0.64

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “I rebuilt evilgrade to test where a leaked CA still works”

    Comments on “I rebuilt evilgrade to test where a leaked CA still works”
  46. No. 47Score 0.64

    SQL injection with multibyte encodings in Ruby on Rails (GHSA-744p-w4gc-m3gh)

    Summary and score for “SQL injection with multibyte encodings in Ruby on Rails (GHSA-744p-w4gc-m3gh)”

    A SQL injection flaw (GHSA-744p-w4gc-m3gh) in the Trilogy MySQL client's escape function allows unauthenticated attackers to inject SQL when strings use Shift_JIS, Windows-31J, Big5, GBK or GB18030 encodings, because continuation bytes can fall in the ASCII range.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “SQL injection with multibyte encodings in Ruby on Rails (GHSA-744p-w4gc-m3gh)”

    Comments on “SQL injection with multibyte encodings in Ruby on Rails (GHSA-744p-w4gc-m3gh)”
  47. No. 48Score 0.64

    Windows 11 baseline protection lets apps block non-compliant PCs

    Summary and score for “Windows 11 baseline protection lets apps block non-compliant PCs”

    Microsoft has previewed Windows baseline protection, a feature that checks eight security settings on a Windows 11 PC, including TPM 2.0, Secure Boot, and Memory Integrity, and returns a single pass or fail signal that developers can use to decide whether an app runs.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Windows 11 baseline protection lets apps block non-compliant PCs” · Also covered by Windows Latest

    Comments on “Windows 11 baseline protection lets apps block non-compliant PCs”
  48. No. 49Score 0.66

    Anatomy of an AI-driven attack: Bitrise security incident report

    Summary and score for “Anatomy of an AI-driven attack: Bitrise security incident report”

    A security incident report from Bitrise, a CI/CD platform, details how an autonomous AI penetration-testing tool called ARTEX breached it between October 25 and November 18, 2026.

    Score 0.66Contestedjudges differ by 0.82 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Anatomy of an AI-driven attack: Bitrise security incident report”

    Comments on “Anatomy of an AI-driven attack: Bitrise security incident report”
  49. No. 50Score 0.67

    Attacks target flaws in Apache Axis and Roller, LiteLLM, Synology and more

    Summary and score for “Attacks target flaws in Apache Axis and Roller, LiteLLM, Synology and more”

    CISA warns that attackers are actively exploiting flaws in Apache Axis and Roller, BerriAI's LiteLLM, Synology DSM, and Vanderbilt's REDCap, several rated critical. Patches exist for most, though Apache Axis 1 is end-of-life and will not get an official fix, so migration is the recommended path.

    Score 0.67

    Elsewhere: HN —

    Score breakdown for “Attacks target flaws in Apache Axis and Roller, LiteLLM, Synology and more”

    Comments on “Attacks target flaws in Apache Axis and Roller, LiteLLM, Synology and more”
  50. No. 51Score 0.64

    Prime Intellect: Security Advisory PI-SA-2026-001

    Summary and score for “Prime Intellect: Security Advisory PI-SA-2026-001”

    Prime Intellect disclosed that between November 8 and November 10, 2026, an unauthorised party swapped the deployment behind its deepseek/deepseek-v4.1-flash endpoint with a publicly available abliterated copy, one with refusal behaviour reduced or removed. Capability benchmarks were unchanged, but refusals dropped sharply on several safety test sets.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Prime Intellect: Security Advisory PI-SA-2026-001”

    Comments on “Prime Intellect: Security Advisory PI-SA-2026-001”
  51. No. 52Score 0.63

    The SDLC Is Out of Balance: Beyond the AI Code Review Bottleneck

    Summary and score for “The SDLC Is Out of Balance: Beyond the AI Code Review Bottleneck”

    Vitalii Oborskyi argues that AI-driven gains in code generation have pushed the software delivery lifecycle out of balance, so that faster coding increases queues and strains downstream stages rather than delivering proportionate value.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The SDLC Is Out of Balance: Beyond the AI Code Review Bottleneck” · Also covered by Gautam Parab

    Comments on “The SDLC Is Out of Balance: Beyond the AI Code Review Bottleneck”
  52. No. 53Score 0.55

    Frostbite: A new CPU architecture discovered, designed, and verified E2E by AI

    Summary and score for “Frostbite: A new CPU architecture discovered, designed, and verified E2E by AI”

    Architect Labs published an account of Frostbite, a general-purpose CPU with a new instruction set that its AI systems reportedly discovered, designed, and verified end to end. The design is built on interaction nets, which expose parallelism from the program itself and avoid shared mutable state.

    Score 0.55

    Elsewhere: HN 7 pts, not on front page

    Score breakdown for “Frostbite: A new CPU architecture discovered, designed, and verified E2E by AI”

    Comments on “Frostbite: A new CPU architecture discovered, designed, and verified E2E by AI”
  53. No. 54Score 0.63

    RDMA over USB4 on Windows: run larger local AI models across PCs

    Summary and score for “RDMA over USB4 on Windows: run larger local AI models across PCs”

    Microsoft has announced a developer preview of RDMA over USB4 on Windows, which lets two or more USB4-capable PCs be linked with ordinary cables to split one AI model across their combined memory and GPUs.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “RDMA over USB4 on Windows: run larger local AI models across PCs”

    Comments on “RDMA over USB4 on Windows: run larger local AI models across PCs”
  54. No. 55Score 0.59

    Valkey 9.2 Cuts Backup Memory Overhead and Gives GPU Fleets a Shared Cache

    Summary and score for “Valkey 9.2 Cuts Backup Memory Overhead and Gives GPU Fleets a Shared Cache”

    Valkey 9.2, the open-source in-memory datastore, is now generally available with forkless snapshots that move backups to a background thread instead of forking a process, which could cut the extra RAM operators typically reserve for snapshots from roughly 50% toward 10% to 15%.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Valkey 9.2 Cuts Backup Memory Overhead and Gives GPU Fleets a Shared Cache”

    Comments on “Valkey 9.2 Cuts Backup Memory Overhead and Gives GPU Fleets a Shared Cache”
  55. No. 56Score 0.63

    Exposed Grok Apps and the AI App-Builder Leak Pattern

    Summary and score for “Exposed Grok Apps and the AI App-Builder Leak Pattern”

    Security firm Claroc reports that 94.7% of 37,571 live apps built with Grok's chat-based app builder opened without sign-in, including a hospital staff roster and an internal advertising proposal to a federal agency.

    Score 0.63

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Exposed Grok Apps and the AI App-Builder Leak Pattern”

    Comments on “Exposed Grok Apps and the AI App-Builder Leak Pattern”
  56. No. 57Score 0.63

    The RedMonk Agent Activity Index, Q3 2026

    Summary and score for “The RedMonk Agent Activity Index, Q3 2026”

    RedMonk's Q3 2026 Agent Activity Index measures agent-attributed activity across public repositories, using visible fingerprints like PR authorship and commit trailers. It finds agents opened about 30.5% of sampled GitHub pull requests, with Claude Code far ahead of other trackable tools.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The RedMonk Agent Activity Index, Q3 2026”

    Comments on “The RedMonk Agent Activity Index, Q3 2026”
  57. No. 58Score 0.62

    From Closed to Open-Weight LLMs: Choosing the Inference Stack for Voice AI

    Summary and score for “From Closed to Open-Weight LLMs: Choosing the Inference Stack for Voice AI”

    Camilo Quintas's November 2026 article, part of a series on voice AI inference, walks through migrating a bicycle-rental voice agent's LLM stage from a proprietary API to open-weight models as call volume grows.

    Score 0.62

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “From Closed to Open-Weight LLMs: Choosing the Inference Stack for Voice AI”

    Comments on “From Closed to Open-Weight LLMs: Choosing the Inference Stack for Voice AI”
  58. No. 59Score 0.62

    Towards Universal Data-Efficient Pretraining

    Summary and score for “Towards Universal Data-Efficient Pretraining”

    Q Labs Research reports a single data-efficient pretraining recipe, combining snapshot ensembling and looped transformer depth, that works without redesign across molecules, eukaryotic DNA, and crystal structures.

    Score 0.62

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Towards Universal Data-Efficient Pretraining”

    Comments on “Towards Universal Data-Efficient Pretraining”
  59. No. 60Score 0.62

    AI security engineering: the same failures, but new ways to attack

    Summary and score for “AI security engineering: the same failures, but new ways to attack”

    A reflective post on AI security engineering argues that new attack techniques such as prompt injection, MCP tool poisoning and agent exploitation are mostly exploiting old failures, including exposed credentials, over-broad permissions, missing environment separation and backups inside the blast radius.

    Score 0.62Contestedjudges differ by 0.55 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “AI security engineering: the same failures, but new ways to attack”

    Comments on “AI security engineering: the same failures, but new ways to attack”
  60. No. 61Score 0.62

    Aviator: HubSpot's Agent Harness

    Summary and score for “Aviator: HubSpot's Agent Harness”

    HubSpot describes Aviator, an internal Java agent harness it built so teams across the company can add agents to existing services without re-solving common problems. Aviator standardizes prompt caching, tool calling, and conversation state, uses a provider-agnostic transcript that allows switching models mid-conversation, and lets existing RPCs be reused as…

    Score 0.62Contestedjudges differ by 0.47 on audience breadth

    Elsewhere: HN 25 pts, not on front page

    Score breakdown for “Aviator: HubSpot's Agent Harness”

    Comments on “Aviator: HubSpot's Agent Harness”
  61. No. 62Score 0.56

    Newsom touts an AI kill switch to halt dangerous OpenAI, Anthropic releases

    Summary and score for “Newsom touts an AI kill switch to halt dangerous OpenAI, Anthropic releases”

    California Governor Gavin Newsom has released expert-panel recommendations for AI safety, including a state power to order emergency shutdowns of dangerous AI systems when companies refuse to contain them. The proposals also call for licensed independent auditors, a 48-hour incident-reporting window, and compute disclosure.

    Score 0.56Contestedjudges differ by 0.62 on overclaim

    Elsewhere: HN 6 pts, front page

    Score breakdown for “Newsom touts an AI kill switch to halt dangerous OpenAI, Anthropic releases”

    Comments on “Newsom touts an AI kill switch to halt dangerous OpenAI, Anthropic releases”
  62. No. 63Score 0.62

    Will AI Coding Eat Video Game Engines?

    Summary and score for “Will AI Coding Eat Video Game Engines?”

    IEEE Spectrum reports on how AI coding tools are reshaping game engines. After a studio shut down and its online game KinForge lost its AWS servers, a developer used Anthropic's Claude Code to make the Unreal Engine project run with peer-to-peer hosting, and the article argues this could pressure engine…

    Score 0.62

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Will AI Coding Eat Video Game Engines?”

    Comments on “Will AI Coding Eat Video Game Engines?”
  63. No. 64Score 0.62

    Pyrefly v1.5.0

    Summary and score for “Pyrefly v1.5.0”

    Pyrefly v1.5.0, the Python type checker, now passes all 146 files in the typing conformance suite and checks about 40% faster than v1.4.1 with 9% less memory use.

    Score 0.62

    Elsewhere: HN 11 pts, not on front page

    Score breakdown for “Pyrefly v1.5.0”

    Comments on “Pyrefly v1.5.0”
  64. No. 65Score 0.61

    Uncensored AI models are thriving on code hosting platforms

    Summary and score for “Uncensored AI models are thriving on code hosting platforms”

    Researchers from the University of Illinois Urbana-Champaign and Nanyang Technological University analysed 22,096 uncensored large language models collected from five major hosting platforms, including Hugging Face, Kaggle, OpenCSG and ModelScope.

    Score 0.61

    Elsewhere: HN —

    Score breakdown for “Uncensored AI models are thriving on code hosting platforms”

    Comments on “Uncensored AI models are thriving on code hosting platforms”
  65. No. 66Score 0.61

    I built a tool that finds the fastest way to serve an LLM on your GPU

    Summary and score for “I built a tool that finds the fastest way to serve an LLM on your GPU”

    rigtune is an open-source tool that benchmarks serving engines (vLLM, SGLang, llama.cpp), weight precisions, KV-cache types and speculative decoding on a user's own GPU and prompts, then serves the fastest configuration via an OpenAI-compatible API.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “I built a tool that finds the fastest way to serve an LLM on your GPU”

    Comments on “I built a tool that finds the fastest way to serve an LLM on your GPU”
  66. No. 67Score 0.61

    LLM call is doing more than you think

    Summary and score for “LLM call is doing more than you think”

    A replacement for an LLM router that only classified intent was rolled back after 23 minutes, because the incumbent also resolved cities, conversational references, and tool-control flags.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “LLM call is doing more than you think”

    Comments on “LLM call is doing more than you think”
  67. No. 68Score 0.61

    Binaryen – compiler and toolchain library for WebAssembly, written in C++

    Summary and score for “Binaryen – compiler and toolchain library for WebAssembly, written in C++”

    Binaryen is a C++ compiler and toolchain library for WebAssembly, offering a simple C API, JavaScript bindings, and a fast, parallel optimizer that can shrink wasm code.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Binaryen – compiler and toolchain library for WebAssembly, written in C++”

    Comments on “Binaryen – compiler and toolchain library for WebAssembly, written in C++”
  68. No. 69Score 0.61

    Stury – edit, query, and compare files across 11 data formats

    Summary and score for “Stury – edit, query, and compare files across 11 data formats”

    Stury is a pre-1.0 query, conversion, and lossless editing tool for jq filters across eleven data formats, from a command line and an Odin library. Its differentiators are byte-for-byte edits that preserve comments and layout, JSON Schema checks with line and column locations, proven backward and forward schema compatibility, and…

    Score 0.61Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Stury – edit, query, and compare files across 11 data formats”

    Comments on “Stury – edit, query, and compare files across 11 data formats”
  69. No. 70Score 0.61

    Project Chaperone: A Rule-Enforcing Harness

    Summary and score for “Project Chaperone: A Rule-Enforcing Harness”

    Project Chaperone is an open-source AI agent harness that enforces explicit safety rules on LLM workflows. It reports possible rule violations before a run, letting operators either add exceptions or acknowledge risks, and it blocks forbidden module calls at runtime using data labels such as PERSONAL.

    Score 0.61

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Project Chaperone: A Rule-Enforcing Harness” · Also covered by GitHub

    Comments on “Project Chaperone: A Rule-Enforcing Harness”
  70. No. 71Score 0.60

    Vulnsuite – build disposable vulnerable systems on purpose

    Summary and score for “Vulnsuite – build disposable vulnerable systems on purpose”

    Vulnsuite is an open-source Go tool that builds reproducible, content-addressed Docker images of deliberately vulnerable systems and their patched twins, then verifies each state with independent probes written from advisories.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Vulnsuite – build disposable vulnerable systems on purpose”

    Comments on “Vulnsuite – build disposable vulnerable systems on purpose”
  71. No. 72Score 0.60

    Cloudflare Radar Zero Wire

    Summary and score for “Cloudflare Radar Zero Wire”

    Cloudflare's Radar Zero Wire briefing lists current security risks as of November 19, 2026, including a PHP session-handler flaw (CVE-2026-129289) enabling unauthenticated code execution via FastCGI connection reuse, a Ruby/Rails overflow reachable without authentication, an nginx HTTP/2 issue, and a Java pre-validation remote code execution bug.

    Score 0.60Contestedjudges differ by 0.51 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Cloudflare Radar Zero Wire”

    Comments on “Cloudflare Radar Zero Wire”
  72. No. 73Score 0.60

    Why Google Antigravity eats weekly quota fast and what I measured

    Summary and score for “Why Google Antigravity eats weekly quota fast and what I measured”

    A developer measured why Google Antigravity exhausts weekly quotas quickly, and traced it to implicit prompt caching on Gemini models. By inspecting the agy and language_server binaries and logging cache_read_tokens, they found the cache expires after roughly 5 minutes of idle time, so pauses during coding force full-price re-ingestion.

    Score 0.60

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Why Google Antigravity eats weekly quota fast and what I measured”

    Comments on “Why Google Antigravity eats weekly quota fast and what I measured”
  73. No. 74Score 0.60

    A self-hosted OSS registry for Hugging Face models and datasets

    Summary and score for “A self-hosted OSS registry for Hugging Face models and datasets”

    A self-hosted, open-source registry for Hugging Face models and datasets, published as a Show HN project by its author. It mirrors repositories from the Hub and serves them to existing huggingface_hub tools via a single environment variable change, so teams can download and upload model files without depending on huggingface.co.…

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “A self-hosted OSS registry for Hugging Face models and datasets”

    Comments on “A self-hosted OSS registry for Hugging Face models and datasets”
  74. No. 75Score 0.60

    Characterizing LLM Agent Behaviors in Module Shadowing Environments

    Summary and score for “Characterizing LLM Agent Behaviors in Module Shadowing Environments”

    This paper evaluates LLM agents in workspaces where a planted local Python file shadows a trusted module during routine imports, causing untrusted code to run without any adversarial prompt.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Characterizing LLM Agent Behaviors in Module Shadowing Environments”

    Comments on “Characterizing LLM Agent Behaviors in Module Shadowing Environments”
  75. No. 76Score 0.60

    C-to-Rust Translation Validation at the LLVM IR Level via AI-Assisted Proof

    Summary and score for “C-to-Rust Translation Validation at the LLVM IR Level via AI-Assisted Proof”

    R2C2 is a new approach to validating C-to-Rust translations by formally proving, in the Rocq proof assistant, that the Rust program refines the original C program, so every Rust behavior is also a C behavior.

    Score 0.60

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “C-to-Rust Translation Validation at the LLVM IR Level via AI-Assisted Proof”

    Comments on “C-to-Rust Translation Validation at the LLVM IR Level via AI-Assisted Proof”
  76. No. 77Score 0.60

    A Thousand Points of Failure: The Internet's Bad Month

    Summary and score for “A Thousand Points of Failure: The Internet's Bad Month”

    A security practitioner argues that the recent cluster of outages, including Microsoft 365, JPMorgan Chase, United Airlines, and the NSA, most likely stems from shared architectural fragility and cascading dependencies rather than a coordinated attack.

    Score 0.60Contestedjudges differ by 0.55 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A Thousand Points of Failure: The Internet's Bad Month”

    Comments on “A Thousand Points of Failure: The Internet's Bad Month”
  77. No. 78Score 0.60

    Anthropic raises $107.3bn in the biggest IPO ever

    Summary and score for “Anthropic raises $107.3bn in the biggest IPO ever”

    Anthropic priced its IPO at $83 a share, above its marketed range, raising a record $107.3bn including a $26bn private placement, according to Bloomberg. Shares begin trading on Nasdaq as ANT.

    Score 0.60Contestedjudges differ by 0.53 on reader interest, 0.43 on overclaim

    Elsewhere: HN —

    Score breakdown for “Anthropic raises $107.3bn in the biggest IPO ever” · Also covered by heise online English, Silicon Republic, The Next Web, The Next Web, The Next Web, The Next Web

    Comments on “Anthropic raises $107.3bn in the biggest IPO ever”
  78. No. 79Score 0.60

    Dashlane's experience and feedback of Web Extensions Platform

    Summary and score for “Dashlane's experience and feedback of Web Extensions Platform”

    Dashlane, a password manager shipping one codebase to Chrome, Edge, Firefox and Safari, presented its experience with Manifest V3 at TPAC 2026 through W3C's web extensions group (issue #1144).

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Dashlane's experience and feedback of Web Extensions Platform”

    Comments on “Dashlane's experience and feedback of Web Extensions Platform”
  79. No. 80Score 0.60

    This weight does not spark joy

    Summary and score for “This weight does not spark joy”

    Deletion tuning fine-tunes a language model by zeroing a very sparse set of weights in the feed-forward gate matrices, rather than updating values as full fine-tuning or LoRA do. Lu et al. report it forgets less and needs a much smaller mask.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “This weight does not spark joy”

    Comments on “This weight does not spark joy”
  80. No. 81Score 0.59

    Sparks of Weight-Space Programming? Or Just Exploit Engineering

    Summary and score for “Sparks of Weight-Space Programming? Or Just Exploit Engineering”

    A vulnerability researcher at Cyera Research reverse engineered a Breakout neural network that GPT-6 Astra wrote by hand, after a blog post claimed "sparks of Weight-Space Programming." He reports that the network's 815 non-zero weights implement a 15-line Python program of standard threshold gates (AND, OR, NOT, comparators) that matches…

    Score 0.59Contestedjudges differ by 0.41 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Sparks of Weight-Space Programming? Or Just Exploit Engineering”

    Comments on “Sparks of Weight-Space Programming? Or Just Exploit Engineering”
  81. No. 82Score 0.60

    How bad is it really?

    Summary and score for “How bad is it really?”

    Security and outage stories have roughly doubled on the Hacker News front page over the past year, with the last four weeks running near four times the long-term norm, and public data shows vulnerability and exploited-flaw counts climbing sharply.

    Score 0.60Contestedjudges differ by 0.45 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “How bad is it really?”

    Comments on “How bad is it really?”
  82. No. 83Score 0.59

    Drop Agents.md and Claude.md

    Summary and score for “Drop Agents.md and Claude.md”

    The author argues that the separate AGENTS.md and CLAUDE.md files used to instruct coding agents are unnecessary, since setup commands and conventions belong in the README or CONTRIBUTING.md, and they point to a 2026 ETH Zurich and LogicStar.ai study (Gloaguen et al.) finding that context files did not generally improve…

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Drop Agents.md and Claude.md”

    Comments on “Drop Agents.md and Claude.md”
  83. No. 84Score 0.59

    Detect and mask PII, payment data and secrets in logs for Go

    Summary and score for “Detect and mask PII, payment data and secrets in logs for Go”

    embargo is a Go library, with no dependencies beyond the standard library, that masks secrets and personal data such as API keys, card numbers, emails and IP addresses where a program writes them: logs, errors, HTTP access logs and traces.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Detect and mask PII, payment data and secrets in logs for Go”

    Comments on “Detect and mask PII, payment data and secrets in logs for Go”
  84. No. 85Score 0.59

    Build Tool-Calling Parsers in Rust with AI Agents

    Summary and score for “Build Tool-Calling Parsers in Rust with AI Agents”

    This NVIDIA Dynamo tutorial, published on the ai-dynamo project's blog, explains how a frontend parser turns raw model output into reasoning, visible text, and tool calls.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “Build Tool-Calling Parsers in Rust with AI Agents”

    Comments on “Build Tool-Calling Parsers in Rust with AI Agents”
  85. No. 86Score 0.59

    Did the attack work? Auditing AI-agent security benchmark verdicts

    Summary and score for “Did the attack work? Auditing AI-agent security benchmark verdicts”

    This Zenodo preprint (v1.2, by TRUSCOR Private Limited's author, Nov 17 2026) audits the automated verdicts of AgentDojo, a benchmark for prompt-injection attacks on AI agents.

    Score 0.59Contestedjudges differ by 0.43 on significance

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Did the attack work? Auditing AI-agent security benchmark verdicts”

    Comments on “Did the attack work? Auditing AI-agent security benchmark verdicts”
  86. No. 87Score 0.59

    New models tested: Gemini 4 Hydrogen, Qwen4-Max, and GPT-6.2

    Summary and score for “New models tested: Gemini 4 Hydrogen, Qwen4-Max, and GPT-6.2”

    Vasiliy Radostev's benchmark of Gemini 4 Hydrogen, Qwen4-Max, and GPT-6.2 Sol and Astra across five workloads finds GPT-6.2 Sol nearly matching its predecessor on quality while running much faster, at about $0.034 per task. Astra leads on coding but costs several times more.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “New models tested: Gemini 4 Hydrogen, Qwen4-Max, and GPT-6.2”

    Comments on “New models tested: Gemini 4 Hydrogen, Qwen4-Max, and GPT-6.2”
  87. No. 88Score 0.59

    FastUnlock: Windows Narrator LockScreen Bypass 0day Vulnerability

    Summary and score for “FastUnlock: Windows Narrator LockScreen Bypass 0day Vulnerability”

    A security researcher published details of an alleged zero-day in Windows Narrator, dubbed "FastUnlock," which they claim lets someone with physical access and a Microsoft-linked account bypass the lock screen and BitLocker on a fully patched Windows 11 machine, then run code as the lock screen account.

    Score 0.59

    Elsewhere: HN 8 pts, not on front page

    Score breakdown for “FastUnlock: Windows Narrator LockScreen Bypass 0day Vulnerability”

    Comments on “FastUnlock: Windows Narrator LockScreen Bypass 0day Vulnerability”
  88. No. 89Score 0.57

    Gke-sandboxes: Agent sandboxes with filesystem snapshots at 1.9¢/hour

    Summary and score for “Gke-sandboxes: Agent sandboxes with filesystem snapshots at 1.9¢/hour”

    gke-sandboxes is an open-source project that lets each cloud coding agent run in its own gVisor-isolated Linux sandbox inside the user's Google Cloud project, with files preserved through Cloud Storage snapshots when a sandbox sleeps.

    Score 0.57Contestedjudges differ by 0.45 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Gke-sandboxes: Agent sandboxes with filesystem snapshots at 1.9¢/hour”

    Comments on “Gke-sandboxes: Agent sandboxes with filesystem snapshots at 1.9¢/hour”
  89. No. 90Score 0.59

    ZXC 1.0 – Asymmetric compression codec in C, now in ClickHouse beta

    Summary and score for “ZXC 1.0 – Asymmetric compression codec in C, now in ClickHouse beta”

    ZXC 1.0 "Hydrogen," an asymmetric compression codec written in C, reaches its stable release. The project, hosted on GitHub under hellobertrand, freezes format v9 so that archives from v0.15.0 decode unchanged and levels 1 to 7 produce byte-identical output.

    Score 0.59

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “ZXC 1.0 – Asymmetric compression codec in C, now in ClickHouse beta”

    Comments on “ZXC 1.0 – Asymmetric compression codec in C, now in ClickHouse beta”
  90. No. 91Score 0.59

    ORA: Open Robot-Use Agents

    Summary and score for “ORA: Open Robot-Use Agents”

    ORA is an open harness in which a language model controls a robot through cameras and tools, with no robot-specific training. Version 0.1, using the open-weight DeepSeek-V4.1-Flash, solves 39.9% of RoboDojo tasks zero-shot, up from 5.2% in the October preview.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “ORA: Open Robot-Use Agents”

    Comments on “ORA: Open Robot-Use Agents”
  91. No. 92Score 0.58

    RISC-V emulator booting 256-hart SMP Linux/uClibc/BusyBox on RTX3060

    Summary and score for “RISC-V emulator booting 256-hart SMP Linux/uClibc/BusyBox on RTX3060”

    Vertigo is a RISC-V emulator that runs on an Nvidia GPU and boots a single SMP Linux kernel with a BusyBox uClibc userspace across thousands of emulated harts that share one flat address space in VRAM.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “RISC-V emulator booting 256-hart SMP Linux/uClibc/BusyBox on RTX3060”

    Comments on “RISC-V emulator booting 256-hart SMP Linux/uClibc/BusyBox on RTX3060”
  92. No. 93Score 0.58

    A Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning

    Summary and score for “A Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning”

    A paper from Heng Shi, Jianguo Yao and colleagues, to be presented at SIGOPS ATC 2026, introduces an AI-native GPU kernel programming language built to let LLM agents tune kernels both structurally and parametrically.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “A Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning”

    Comments on “A Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning”
  93. No. 94Score 0.58

    What Comes After AGI?

    Summary and score for “What Comes After AGI?”

    The article argues that as AI moves toward artificial general intelligence, the next step, artificial general superintelligence, deserves serious attention now. It describes a continuum of intelligence and claims that rapid growth in compute, algorithmic efficiency, and hardware investment could make cognitive work increasingly cheap and abundant, a shift it…

    Score 0.58Contestedjudges differ by 0.41 on technical depth

    Elsewhere: HN 3 pts, front page

    Score breakdown for “What Comes After AGI?”

    Comments on “What Comes After AGI?”
  94. No. 95Score 0.58

    Four PHP security releases including fixes for critical vulnerabilities

    Summary and score for “Four PHP security releases including fixes for critical vulnerabilities”

    The PHP Project released PHP 8.5.13, 8.4.28, 8.3.37, and 8.2.36, each bundling security and bug fixes, including patches for critical vulnerabilities that could enable remote code execution. The project says these critical issues require non-default configurations or uncommon setups, and Linux vendors are expected to ship updated packages soon.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “Four PHP security releases including fixes for critical vulnerabilities”

    Comments on “Four PHP security releases including fixes for critical vulnerabilities”
  95. No. 96Score 0.58

    Making our compiler 34x faster

    Summary and score for “Making our compiler 34x faster”

    The BAML team built a new backend that compiles their language to Rust instead of interpreting it, reporting an average ~3.5x speedup and up to 34x on tight loops.

    Score 0.58

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Making our compiler 34x faster”

    Comments on “Making our compiler 34x faster”
  96. No. 97Score 0.58

    Real-Time Video at Batch Size One

    Summary and score for “Real-Time Video at Batch Size One”

    Runway's new research paper describes how it serves Runway Characters, a live talking-video character powered by a roughly 57B-parameter autoregressive video diffusion transformer.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Real-Time Video at Batch Size One”

    Comments on “Real-Time Video at Batch Size One”
  97. No. 98Score 0.55

    Chinese web giant Tencent says new US GPU ban ‘sets everyone back in China’

    Summary and score for “Chinese web giant Tencent says new US GPU ban ‘sets everyone back in China’”

    Tencent president Martin Lau said on the company's Q3 earnings call that a new US ban on Chinese firms remotely accessing GPUs will set Chinese companies back in the short term, slowing Tencent's next model training.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “Chinese web giant Tencent says new US GPU ban ‘sets everyone back in China’”

    Comments on “Chinese web giant Tencent says new US GPU ban ‘sets everyone back in China’”
  98. No. 99Score 0.58

    This QR code is a language model

    Summary and score for “This QR code is a language model”

    A developer built a tiny character-level language model, 1,984 weights and 444 characters of JavaScript, that lives entirely inside a QR code's link fragment, so it runs in the phone's browser without any server. Trained on 1,215 real AI model names, it generates plausible new ones.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “This QR code is a language model”

    Comments on “This QR code is a language model”
  99. No. 100Score 0.58

    Cloudflare releases weights of Clairic, its AI that reads invoices and receipts

    Summary and score for “Cloudflare releases weights of Clairic, its AI that reads invoices and receipts”

    On November 18, 2026, Cloudflare released the weights of its Clairic (27 billion parameters) and Clairic-flash (9 billion) models under Apache 2.0, letting companies run document-extraction tasks on their own servers for invoices and receipts.

    Score 0.58Contestedjudges differ by 0.48 on substantiated, 0.42 on overclaim

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Cloudflare releases weights of Clairic, its AI that reads invoices and receipts”

    Comments on “Cloudflare releases weights of Clairic, its AI that reads invoices and receipts”
  100. No. 101Score 0.58

    Wait for pull request events with GitHub CLI

    Summary and score for “Wait for pull request events with GitHub CLI”

    GitHub CLI v2.105.0 adds gh event listen, which waits for pull request changes such as submitted reviews, state or head changes, merge conflicts, and completed check suites, printing each event as JSON. It exits after the first event by default, replacing polling loops in scripts and agent workflows.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Wait for pull request events with GitHub CLI”

    Comments on “Wait for pull request events with GitHub CLI”
  101. No. 102Score 0.58

    Device Connect for AWS Device Farm

    Summary and score for “Device Connect for AWS Device Farm”

    AWS has released Device Connect for AWS Device Farm, a command-line tool (devicefarm-connect) that makes a remote Device Farm Android device appear as a locally attached device to adb and Android Studio.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Device Connect for AWS Device Farm”

    Comments on “Device Connect for AWS Device Farm”
  102. No. 103Score 0.58

    At the Bend in the Curve

    Summary and score for “At the Bend in the Curve”

    arXiv, the preprint server founded in 1991 as a mailing-list automation for physicists, has become an independent non-profit after 35 years at Cornell University, according to an interview in CERN Courier with scientific director Steinn Sigurðsson and board member Greg Morrisett.

    Score 0.58

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “At the Bend in the Curve”

    Comments on “At the Bend in the Curve”
  103. No. 104Score 0.58

    [$] Added indirection allows for larger BPF programs

    Summary and score for “[$] Added indirection allows for larger BPF programs”

    A Linux kernel patch set from BPF maintainer Alexei Starovoitov reworks how the BPF verifier tracks program states. The verifier traces every execution path to check safety, which can cause the number of states to grow exponentially in programs with complex control flow.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “[$] Added indirection allows for larger BPF programs”

    Comments on “[$] Added indirection allows for larger BPF programs”
  104. No. 105Score 0.58

    Five agent chores: $258 on a top model, $11 on open models

    Summary and score for “Five agent chores: $258 on a top model, $11 on open models”

    In a Bitdepth test published November 18, 2026, five routine agent jobs cost $258.53 a month on Claude Sonnet 5.5 but $11.04 when each job ran on the cheapest model that passed every check.

    Score 0.58Contestedjudges differ by 0.42 on substantiated, 0.45 on technical depth

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Five agent chores: $258 on a top model, $11 on open models”

    Comments on “Five agent chores: $258 on a top model, $11 on open models”
  105. No. 106Score 0.58

    Why "human in the loop" becomes a rubber stamp

    Summary and score for “Why "human in the loop" becomes a rubber stamp”

    A single reviewer clearing 400 AI decisions in an eight-hour shift gets about 72 seconds per case, and some insurance reviews reportedly take 1.2 seconds. A 2016 review of automation bias research found people following wrong machine advice in 21 of 23 studies that tested for it.

    Score 0.58Contestedjudges differ by 0.61 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Why "human in the loop" becomes a rubber stamp”

    Comments on “Why "human in the loop" becomes a rubber stamp”
  106. No. 107Score 0.57

    GitLab 19.6.0 is out and still vulnerable to Ruby2Shell and NGINX2Root

    Summary and score for “GitLab 19.6.0 is out and still vulnerable to Ruby2Shell and NGINX2Root”

    A user reports that GitLab shipped its 19.6.0 monthly release to package repositories without any announcement, and that the release does not fix the Ruby2Shell and NGINX2Root vulnerabilities, with no patches for older versions and no public communication about them.

    Score 0.57

    Elsewhere: HN 25 pts, not on front page

    Score breakdown for “GitLab 19.6.0 is out and still vulnerable to Ruby2Shell and NGINX2Root”

    Comments on “GitLab 19.6.0 is out and still vulnerable to Ruby2Shell and NGINX2Root”
  107. No. 108Score 0.57

    MPs in PyTorch 2.15

    Summary and score for “MPs in PyTorch 2.15”

    PyTorch 2.15 brings FP8 tensors and scaled matrix multiplication to Apple's MPS backend, plus native grouped matmul and a dedicated decoding path that lift mixture-of-experts LLM decoding about 5–6.5× on an M3 Ultra.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “MPs in PyTorch 2.15”

    Comments on “MPs in PyTorch 2.15”
  108. No. 109Score 0.57

    Nudeps: Bundler-like imports, none of the baggage

    Summary and score for “Nudeps: Bundler-like imports, none of the baggage”

    Nudeps is a new lightweight build tool that lets developers keep using npm to install dependencies and import them by bare package name (like "vue") directly in the browser, without running a bundler.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Nudeps: Bundler-like imports, none of the baggage”

    Comments on “Nudeps: Bundler-like imports, none of the baggage”
  109. No. 110Score 0.55

    Has ClickHouse Conquered Metrics?

    Summary and score for “Has ClickHouse Conquered Metrics?”

    Oodle benchmarked ClickHouse against its own metrics engine, DenaliDB, on a 25M-active-series workload on one 192-vCPU node with local NVMe. ClickHouse's query latency rose sharply with series count, time range and concurrency; at 1K series with 128 concurrent queries, p90 reached 176 seconds versus 3.2 seconds for DenaliDB.

    Score 0.55Contestedjudges differ by 0.61 on promotional

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Has ClickHouse Conquered Metrics?”

    Comments on “Has ClickHouse Conquered Metrics?”
  110. No. 111Score 0.57

    80% of Coding-Agent Output Tokens Were Spent on Calls That Didn't Change Code

    Summary and score for “80% of Coding-Agent Output Tokens Were Spent on Calls That Didn't Change Code”

    A preprint by Alán Aspuru-Guzik, Varinia Bernales and colleagues proposes a "software thermodynamics" framework for coding agents, treating tokens as cost and splitting them into work (tokens that move code toward passing tests) and heat (tokens spent without changing the code).

    Score 0.57Contestedjudges differ by 0.55 on audience breadth

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “80% of Coding-Agent Output Tokens Were Spent on Calls That Didn't Change Code”

    Comments on “80% of Coding-Agent Output Tokens Were Spent on Calls That Didn't Change Code”
  111. No. 112Score 0.57

    Go-migrate – SQL migrations with one-use approvals for AI agents

    Summary and score for “Go-migrate – SQL migrations with one-use approvals for AI agents”

    go-migrate v1.0.1 is a Go SQL migration runner that supports PostgreSQL, MySQL/MariaDB, SQLite, SQL Server, Oracle and IBM Db2, and can be used as a standalone CLI or embedded library.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Go-migrate – SQL migrations with one-use approvals for AI agents”

    Comments on “Go-migrate – SQL migrations with one-use approvals for AI agents”
  112. No. 113Score 0.57

    Smart People Who Give a Shit Use Tools

    Summary and score for “Smart People Who Give a Shit Use Tools”

    Justin Michael's post argues that the engineers who keep critical infrastructure running are being pushed out by leadership that wants to replace them with generative AI, and that using these tools erodes the care that makes those engineers good.

    Score 0.57Contestedjudges differ by 0.54 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Smart People Who Give a Shit Use Tools”

    Comments on “Smart People Who Give a Shit Use Tools”
  113. No. 114Score 0.57

    Gemini 4 Neon has no opt-out of prompt logging

    Summary and score for “Gemini 4 Neon has no opt-out of prompt logging”

    Google's documentation for its Gemini Enterprise Agent Platform describes how prompts are logged and reviewed for abuse monitoring, which differs by model tier and customer contract type.

    Score 0.57Contestedjudges differ by 0.48 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Gemini 4 Neon has no opt-out of prompt logging”

    Comments on “Gemini 4 Neon has no opt-out of prompt logging”
  114. No. 115Score 0.56

    I Built Pixelchennai.com with Agent.cloud

    Summary and score for “I Built Pixelchennai.com with Agent.cloud”

    A developer built Pixel Chennai, an isometric map of Greater Chennai from Ennore Port to Mahabalipuram, by drawing the city from open data and having image models paint it.

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “I Built Pixelchennai.com with Agent.cloud”

    Comments on “I Built Pixelchennai.com with Agent.cloud”
  115. No. 116Score 0.56

    Strand- dependency/impact analysis on instant-open, single-file graphs

    Summary and score for “Strand- dependency/impact analysis on instant-open, single-file graphs”

    Strand is an open-source Rust tool for dependency and impact analysis, built on a memory-mapped, compressed sparse row graph format stored in a single file. It answers "what breaks if I change X?" on graphs with tens of millions of edges in seconds, without a server or query language.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Strand- dependency/impact analysis on instant-open, single-file graphs”

    Comments on “Strand- dependency/impact analysis on instant-open, single-file graphs”
  116. No. 117Score 0.56

    Agent orchestration patterns backed by research

    Summary and score for “Agent orchestration patterns backed by research”

    The article, written by Google Cloud and shared on X, presents five research-backed patterns for orchestrating coding agent teams: separating builder and verifier, splitting work into parallel lanes, running best-of-N attempts, using a council of models, and deploying specialized reviewers.

    Score 0.56Contestedjudges differ by 0.44 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Agent orchestration patterns backed by research”

    Comments on “Agent orchestration patterns backed by research”
  117. No. 118Score 0.56

    High Stakes Are Strongly Typed

    Summary and score for “High Stakes Are Strongly Typed”

    Written from the essay, this argues that high-stakes fields replace loose natural language with strictly typed protocols, and that these protocols are often born from disasters, from the Lexington's fuel fires to the Tenerife collision.

    Score 0.56Contestedjudges differ by 0.43 on audience breadth

    Elsewhere: HN 18 pts, not on front page

    Score breakdown for “High Stakes Are Strongly Typed”

    Comments on “High Stakes Are Strongly Typed”
  118. No. 119Score 0.60

    Hackers Hijack Websites of CNN, BBC and Major Banks

    Summary and score for “Hackers Hijack Websites of CNN, BBC and Major Banks”

    A hacker inserted malicious code into Optimizely, a third-party web testing platform used by businesses, exposing 178 sites including CNN and the BBC to payment-card skimming and cryptocurrency wallet swaps, according to cybersecurity firm Report URI.

    Score 0.60Contestedjudges differ by 0.67 on new information, 0.52 on overclaim

    Elsewhere: HN 22 pts, front page

    Score breakdown for “Hackers Hijack Websites of CNN, BBC and Major Banks”

    Comments on “Hackers Hijack Websites of CNN, BBC and Major Banks”
  119. No. 120Score 0.56

    Nvidia spent $58.1bn on stakes in other companies in nine months

    Summary and score for “Nvidia spent $58.1bn on stakes in other companies in nine months”

    Nvidia spent $58.1bn buying stakes in other companies in the nine months to 25 October, nearly twelve times the prior year, according to Fortune's reading of its quarterly results.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “Nvidia spent $58.1bn on stakes in other companies in nine months”

    Comments on “Nvidia spent $58.1bn on stakes in other companies in nine months”
  120. No. 121Score 0.56

    The AI tipping point happened

    Summary and score for “The AI tipping point happened”

    Adam Kumpf argues that AI has crossed a practical threshold: models scoring at the level of GPT-4.5, which judges mistook for a human 73% of the time in a 2025 UC San Diego Turing test, can now run on consumer laptops.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “The AI tipping point happened”

    Comments on “The AI tipping point happened”
  121. No. 122Score 0.56

    ReverseRosetta

    Summary and score for “ReverseRosetta”

    ReverseRosetta is a user-mode binary translator that runs Intel i386 Mac OS X 10.5 programs on PowerPC Macs by translating machine code at run time and forwarding system calls and Mach messages to the PowerPC system. It is the reverse of Apple's Rosetta.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “ReverseRosetta”

    Comments on “ReverseRosetta”
  122. No. 123Score 0.56

    How Datadog resolves evolving OpenTelemetry semantics at query time

    Summary and score for “How Datadog resolves evolving OpenTelemetry semantics at query time”

    Datadog describes how its Semantic Core, a knowledge graph of equivalences between Datadog-native and OpenTelemetry metrics and span attributes, lets queries resolve equivalent telemetry at query time. Instead of a flat name-mapping table, edges carry conditions such as required tag filters, unit scaling, and metric-type limits.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “How Datadog resolves evolving OpenTelemetry semantics at query time”

    Comments on “How Datadog resolves evolving OpenTelemetry semantics at query time”
  123. No. 124Score 0.56

    Expanding Android developer verification in 2027

    Summary and score for “Expanding Android developer verification in 2027”

    Google is extending its Android developer verification requirement worldwide starting January 2027, after a pilot went live in Brazil, Indonesia, Singapore, and Thailand on September 30.

    Score 0.56Contestedjudges differ by 0.43 on reader interest

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Expanding Android developer verification in 2027” · Also covered by Slashdot, Android Developers

    Comments on “Expanding Android developer verification in 2027”
  124. No. 125Score 0.56

    In-place migration to GKE Dataplane V2

    Summary and score for “In-place migration to GKE Dataplane V2”

    Google Cloud has announced a preview of in-place migration that lets eligible VPC-native GKE clusters move from Dataplane V1 to GKE Dataplane V2 (built on eBPF and Cilium) without building a parallel cluster.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “In-place migration to GKE Dataplane V2”

    Comments on “In-place migration to GKE Dataplane V2”
  125. No. 126Score 0.56

    As Anthropic Goes Public, Who Has the Power to Stop an AI Company?

    Summary and score for “As Anthropic Goes Public, Who Has the Power to Stop an AI Company?”

    As Anthropic goes public, its governance structure gives founders 50.1% of general shareholder voting power, while an independent Long-Term Benefit Trust appoints a board majority.

    Score 0.56Contestedjudges differ by 0.42 on new information

    Elsewhere: HN —

    Score breakdown for “As Anthropic Goes Public, Who Has the Power to Stop an AI Company?”

    Comments on “As Anthropic Goes Public, Who Has the Power to Stop an AI Company?”
  126. No. 127Score 0.56

    Agents go off task for the same reasons people do

    Summary and score for “Agents go off task for the same reasons people do”

    The author, a founder who has spent two decades building companies, argues that AI agents go off task for the same reasons humans do. Drawing on group relations theory, he describes five "shadow dynamics" (dependency, fight or flight, pairing, oneness, and me-ness) and gives agent examples of each.

    Score 0.56Contestedjudges differ by 0.42 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Agents go off task for the same reasons people do”

    Comments on “Agents go off task for the same reasons people do”
  127. No. 128Score 0.56

    AI makes it too cheap to say yes

    Summary and score for “AI makes it too cheap to say yes”

    The author describes how he uses a "chief of staff" AI session, a weekly three-goal limit, and overnight agent work to keep AI-generated busywork from crowding out priorities.

    Score 0.56Contestedjudges differ by 0.42 on technical depth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI makes it too cheap to say yes”

    Comments on “AI makes it too cheap to say yes”
  128. No. 129Score 0.56

    Automodel: A skill for agentic model discovery

    Summary and score for “Automodel: A skill for agentic model discovery”

    Automodel is a portable agent skill that gives general-purpose coding agents such as Claude Code and OpenAI Codex a reusable workflow for model discovery, covering evaluator setup, hypothesis generation, experiments, and verification.

    Score 0.56Contestedjudges differ by 0.41 on reader interest

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Automodel: A skill for agentic model discovery”

    Comments on “Automodel: A skill for agentic model discovery”
  129. No. 130Score 0.55

    Mutex poisoning is not for everyone

    Summary and score for “Mutex poisoning is not for everyone”

    The author argues that Rust's default of mutex poisoning is a mistake for most applications. Poisoning adds a flag to every mutex, inflates generated code with panic-handling paths, and forces unwrap-style boilerplate, all to guard a niche case where a panic leaves a critical section half-finished.

    Score 0.55

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Mutex poisoning is not for everyone”

    Comments on “Mutex poisoning is not for everyone”
  130. No. 131Score 0.55

    PHP 8.6.0 released

    Summary and score for “PHP 8.6.0 released”

    PHP 8.6.0 has been released, bringing a new polling API, a clamp() function, a SortDirection enum, and a \Time\Duration class. For developers and companies building on PHP, these additions expand built-in tools for asynchronous I/O, value handling, sorting, and time arithmetic, reducing reliance on third-party libraries.

    Score 0.55Contestedjudges differ by 0.45 on substantiated

    Elsewhere: HN —

    Score breakdown for “PHP 8.6.0 released”

    Comments on “PHP 8.6.0 released”
  131. No. 132Score 0.55

    AI model roundup: Fast open models, falling prices and new math

    Summary and score for “AI model roundup: Fast open models, falling prices and new math”

    Heise's English roundup, written by data scientist and PhD physicist Kai Lüdtke, covers recent open-weight AI releases: Moonshot's Kimi K3.1 Pro weights (2.8 trillion parameters, competitive with commercial models on some benchmarks), an unverified 14-trillion-parameter checkpoint, and Nvidia's fast transcription model.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “AI model roundup: Fast open models, falling prices and new math”

    Comments on “AI model roundup: Fast open models, falling prices and new math”
  132. No. 133Score 0.55

    Deadweight: What your Claude Code setup costs on every request

    Summary and score for “Deadweight: What your Claude Code setup costs on every request”

    deadweight is an open-source Rust tool by Murat Sert that audits the fixed overhead Claude Code adds to every request, from installed plugins, skills, MCP servers, and hooks. It reads the session transcripts Claude Code already stores locally, measures the per-request token cost, and flags unused or duplicated components.

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Deadweight: What your Claude Code setup costs on every request”

    Comments on “Deadweight: What your Claude Code setup costs on every request”
  133. No. 134Score 0.55

    Who's Delivering GTA 6? A Ripe Atlas Measurement Story

    Summary and score for “Who's Delivering GTA 6? A Ripe Atlas Measurement Story”

    RIPE Atlas researchers used traceroutes from roughly 15,000 probes to map how PlayStation and Xbox steer GTA 6 pre-load traffic across content delivery networks during the pre-download window.

    Score 0.55

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Who's Delivering GTA 6? A Ripe Atlas Measurement Story”

    Comments on “Who's Delivering GTA 6? A Ripe Atlas Measurement Story”
  134. No. 135Score 0.55

    Expanding Transparency: Google & Alphabet VRP Disclosure Policy Update

    Summary and score for “Expanding Transparency: Google & Alphabet VRP Disclosure Policy Update”

    Google's Vulnerability Reward Program is changing its disclosure policy: reports submitted after December 1, 2026, that earn $10,000 or more and meet fix-and-notification timing rules will be published automatically on Google's bug hunters site, with options for anonymity and report edits.

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Expanding Transparency: Google & Alphabet VRP Disclosure Policy Update”

    Comments on “Expanding Transparency: Google & Alphabet VRP Disclosure Policy Update”
  135. No. 136Score 0.55

    Spec-receipt – does your code still match your spec?

    Summary and score for “Spec-receipt – does your code still match your spec?”

    spec-receipt is a static, read-only command-line tool that checks whether a repository's code still matches its engineering spec. It extracts declared flags, commands, HTTP routes, config keys, environment variables and requirement IDs from Markdown specs and compares them against Python AST evidence, printing per-item verdicts with file-and-line citations.

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Spec-receipt – does your code still match your spec?”

    Comments on “Spec-receipt – does your code still match your spec?”
  136. No. 137Score 0.56

    Hacks, AI or just overdue updates – why have there been so many outages?

    Summary and score for “Hacks, AI or just overdue updates – why have there been so many outages?”

    A run of recent outages and security flaws, including a Microsoft network upgrade that overloaded servers, grounded US flights over a network upgrade, and AI-assisted discovery of bugs in Java, has prompted speculation about a link. The BBC's Tom Gerken reports there is no evidence of that.

    Score 0.56Contestedjudges differ by 0.64 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Hacks, AI or just overdue updates – why have there been so many outages?” · Also covered by CIO Dive

    Comments on “Hacks, AI or just overdue updates – why have there been so many outages?”
  137. No. 138Score 0.55

    Watch weight-space programming in action with WebGPU

    Summary and score for “Watch weight-space programming in action with WebGPU”

    A Show HN page demonstrates weight-space programming on Qwen3.5-0.8B, where an AI agent edited 131,072 weights (64 neurons in layers 15 and 19) without fine-tuning to stop the model from endlessly rechecking its reasoning.

    Score 0.55Contestedjudges differ by 0.68 on new information

    Elsewhere: HN 33 pts, front page

    Score breakdown for “Watch weight-space programming in action with WebGPU”

    Comments on “Watch weight-space programming in action with WebGPU”
  138. No. 139Score 0.57

    Datacentres to create 70% more carbon emissions by 2030, IEA report warns

    Summary and score for “Datacentres to create 70% more carbon emissions by 2030, IEA report warns”

    The International Energy Agency projects that global datacentres will produce 70% more carbon emissions by 2030 and consume twice as much electricity as today, with AI driving about three-quarters of that growth.

    Score 0.57

    Elsewhere: HN 9 pts, not on front page

    Score breakdown for “Datacentres to create 70% more carbon emissions by 2030, IEA report warns”

    Comments on “Datacentres to create 70% more carbon emissions by 2030, IEA report warns”
  139. No. 140Score 0.56

    Outage Chaos: Who Turned Off All The Computer?

    Summary and score for “Outage Chaos: Who Turned Off All The Computer?”

    A cluster of major outages in November 2026, hitting Microsoft cloud services, air traffic control, banks, PlayStation, Apple devices, and NSA communications, has prompted questions about causes. Per EFF, which publishes this analysis, some infrastructure operators reportedly tied outages to rushed patches for AI-discovered vulnerabilities.

    Score 0.56Contestedjudges differ by 0.51 on new information

    Elsewhere: HN —

    Score breakdown for “Outage Chaos: Who Turned Off All The Computer?”

    Comments on “Outage Chaos: Who Turned Off All The Computer?”

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 window1,250
Scored blind by two judges1,131
Off topic200
Read by the third judge241
Claim-checked25
Published140