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 111 stories · 11 hidden by your filters

  1. No. 1Score 0.79

    Optimizely Compromised

    Summary and score for “Optimizely Compromised”

    Researcher Amirul Cyber reports that attacker JavaScript in Optimizely-served bundles loads a stealer, track.js, on 15 captured page loads, including CNN, Zoom, BBC, Salesforce and GitLab. It swaps crypto wallet addresses, skims card data and exfiltrates cookies and screenshots. Twelve loads came through Optimizely bundles; the rest via an ad wrapper, a first-party bundle and a Jscrambler script. Two sites blocked it with CSP. The findings are unsourced and not independently confirmed. Teams using tag managers should search served bundles for the listed hosts and monitor CSP violations.

    Score 0.79Claim check: 8 no source given

    Elsewhere: HN 187 pts, front page

    Score breakdown and claims for “Optimizely Compromised”

    Comments on “Optimizely Compromised”
  2. No. 2Score 0.79

    I posted in "Who wants to be hired?" and a recruiter sent me a RAT and a stealer

    Summary and score for “I posted in "Who wants to be hired?" and a recruiter sent me a RAT and a stealer”

    Pavel Savinov analyzed a fake recruiter's take-home test, a VS Code extension (huntrole.assessment-tool) that shows candidates harmless code while the server sends a different script, about 8.7 kB, for the editor to execute. He read the files without running them. The malware finds its server through zero-value Ethereum transactions, which exposes four months of server changes on a public ledger. Links to North Korean-attributed families and to other campaigns rest on his analysis, and several of his claims could not be independently checked.

    Score 0.79Claim check: 1 matches source · 3 no source given · 4 couldn’t check

    Elsewhere: HN 39 pts, not on front page

    Score breakdown and claims for “I posted in "Who wants to be hired?" and a recruiter sent me a RAT and a stealer”

    Comments on “I posted in "Who wants to be hired?" and a recruiter sent me a RAT and a stealer”
  3. No. 3Score 0.78

    NoisyTCP

    Summary and score for “NoisyTCP”

    Hyunwoo Kim disclosed NoisyTCP (CVE-2026-99288), a Linux TCP race present since a 2015 commit and fixed on 2026-11-03. It needs packets handled on separate RX queues or CPUs. The easier path ends in a NULL listener dereference and a remote kernel panic; a use-after-free path is described as much harder. Kim reports 5 seconds to 2 minutes on VMs but about 14 hours on a quiet 8-queue bare-metal host. These figures are his own. Mitigations: disable SYN cookies per network namespace, or use a single RX queue.

    Score 0.78Contestedjudges differ by 0.84 on new informationClaim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “NoisyTCP”

    Comments on “NoisyTCP”
  4. No. 4Score 0.74

    What Happens to Work After a Timeout?

    Summary and score for “What Happens to Work After a Timeout?”

    Emad Khair at octri.dev shows that a timeout limits how long a caller waits, not how long the work runs. In his test, a 350 ms operation behind a 100 ms deadline returned an error on time in Node.js and Python, then finished anyway. Go contexts and Tokio futures cancel only what observes them, and blocking threads keep running. Khair recommends propagating one deadline, bounding concurrency, using idempotency keys for retries, and tracking work that outlives its caller. His capacity figures are idealized arithmetic.

    Score 0.74Claim check: 5 matches source · 3 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “What Happens to Work After a Timeout?”

    Comments on “What Happens to Work After a Timeout?”
  5. No. 5Score 0.75

    Exploiting Replay Protected Monotonic Counters to Brute-Force Windows Hello Pins

    Summary and score for “Exploiting Replay Protected Monotonic Counters to Brute-Force Windows Hello Pins”

    TU Berlin researchers (ACM CCS '26) show that firmware TPM lockout state can be rolled back on x86 laptops with a low-cost SPI interposer. Winbond flash leaks its RPMC HMAC key under voltage glitching, and AMD's NV controller does not verify counter increments. The authors report recovering a 4-digit Windows Hello PIN in 40 hours instead of 69 days, and a 6-digit BitLocker PIN in 20 days instead of 19 years.

    Score 0.75Claim check: 7 no source given · 1 couldn’t check

    Elsewhere: HN 7 pts, not on front page

    Score breakdown and claims for “Exploiting Replay Protected Monotonic Counters to Brute-Force Windows Hello Pins”

    Comments on “Exploiting Replay Protected Monotonic Counters to Brute-Force Windows Hello Pins”
  6. No. 6Score 0.71

    Rust 1.100 stabilizes Never Type after 10 years; Cargo holds new releases

    Summary and score for “Rust 1.100 stabilizes Never Type after 10 years; Cargo holds new releases”

    Rust 1.100 stabilizes the Never Type ! and core parts of the Allocator API, so per-collection allocators for Vec and Box no longer need Nightly. Cargo gains a configurable minimum publish age, which skips freshly released crate versions not already in Cargo.lock, and it only works where the registry records publish times. In editions before 2024 the never-type fallback changes from () to !, which can break code that ignores a generic result. Heise advises checking CI logs for the related lints before upgrading.

    Score 0.71Contestedjudges differ by 0.78 on new informationClaim check: 4 matches source · 1 no source given · 3 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “Rust 1.100 stabilizes Never Type after 10 years; Cargo holds new releases”

    Comments on “Rust 1.100 stabilizes Never Type after 10 years; Cargo holds new releases”
  7. No. 7Score 0.76

    Optimizely Compromised, Served Malware on Zoom, GitLab

    Summary and score for “Optimizely Compromised, Served Malware on Zoom, GitLab”

    Report URI's Scott Helme reports that on 16 November, code written into the Project JavaScript field of Optimizely Web customer snippets served a crypto-address swapper, card skimmer and iOS exploit loaders (Coruna, DarkSword) from 188 snippets across 178 organisations, including Zoom, the BBC and GitLab. About $7,040 had been stolen by 19:25 UTC, and one snippet was still infected at 21:52. How the attacker gained access is unknown. Figures come from Report URI's own crawl.

    Score 0.76Contestedjudges differ by 0.64 on promotionalClaim check: 8 no source given

    Elsewhere: HN 225 pts, front page

    Score breakdown and claims for “Optimizely Compromised, Served Malware on Zoom, GitLab”

    Comments on “Optimizely Compromised, Served Malware on Zoom, GitLab”
  8. No. 8Score 0.70

    Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use

    Summary and score for “Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use”

    Security researcher Xusheng Li reverse engineered MissionControlAggregator.dll in Windows 11 and found it flags which coding agents (Codex, Claude, Copilot, Cursor) and developer tools (Python, GCC, dotnet) have run, using only executable names. The result is a bitmap sent as a Segments.SubCohorts event, queued for upload when optional diagnostics are on, which was the default in his US setup. The feature was active on his ARM64 VM but disabled in the public ISO baseline. Matching is by file name, and the findings come from one researcher's tests.

    Score 0.70Claim check: 8 no source given

    Elsewhere: HN 415 pts, front page

    Score breakdown and claims for “Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use”

    Comments on “Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use”
  9. No. 9Score 0.69

    Top Eval Papers of 2026

    Summary and score for “Top Eval Papers of 2026”

    Elvis Saravia's roundup of ten 2026 eval papers argues that weak benchmarks mislead product decisions, and that agents make grading harder. Findings include nearly half of 60 benchmarks being saturated, SkillsBench raising pass rates from 33.9% to 50.5% in a paired with/without comparison, and Claw-Eval showing that final-output grading missed 44% of safety violations. Nine LLM judges carried roughly two independent votes. The Epoch AI and OpenAI audit figures could not be verified. The advice: run private, paired evals on your own tasks.

    Score 0.69Claim check: 6 matches source · 2 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Top Eval Papers of 2026”

    Comments on “Top Eval Papers of 2026”
  10. No. 10Score 0.68

    Enabling the parallel frontend on nightly

    Summary and score for “Enabling the parallel frontend on nightly”

    The Rust project has turned on the parallel compiler frontend by default on nightly since nightly-2026-11-14, using 2 threads, to find bugs before stabilizing it. In the project's own benchmarks, single-crate cargo check builds were 44% faster with 8 threads, and some crates nearly 3x. Checking the whole cargo crate graph on 8 cores improved only about 15%, since cores were already busy. Diagnostic order may vary, and some binaries may not be reproducible. The figures come from the Rust team and were not independently verified.

    Score 0.68Claim check: 6 no source given · 2 couldn’t check

    Elsewhere: HN 10 pts, not on front page

    Score breakdown and claims for “Enabling the parallel frontend on nightly”

    Comments on “Enabling the parallel frontend on nightly”
  11. No. 11Score 0.68

    Kava Lend guarded its interest math against zero. The attacker brought two

    Summary and score for “Kava Lend guarded its interest math against zero. The attacker brought two”

    Kava Lend lost most of its main KAVA pool on November 14 to a flaw in its interest math, according to a post-mortem by nefariousplan.com. The BUSD market's net supply was pushed to two base units, so the supply index exploded and inflated the attacker's collateral. Its guard covers only zero. The author reports the fix is a trillion-dollar minimum loan setting, not a code change, and similar empty-market designs elsewhere rely on caps. The dollar figures and an earlier ATOM shortfall are the author's own, uncorroborated readings.

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

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Kava Lend guarded its interest math against zero. The attacker brought two”

    Comments on “Kava Lend guarded its interest math against zero. The attacker brought two”
  12. No. 12Score 0.74

    Atlassian Confluence Data Center Exploited via Apache Axis RCE (CVE-2026-90436)

    Summary and score for “Atlassian Confluence Data Center Exploited via Apache Axis RCE (CVE-2026-90436)”

    Rapid7 reports active exploitation of CVE-2026-90436, a critical unauthenticated remote code execution flaw (CVSS 9.8) in the end-of-life Apache Axis 1.x SOAP library, which Apache will not fix. Rapid7 says it detected attacks on Confluence Data Center on November 13, and N-able reported similar attacks on N-central. Both products are vulnerable only with non-default settings: Confluence's Remote API (fixed in 9.2.28) or N-central's Legacy API. Most claims come from Rapid7 and the vendors and could not be independently checked. Teams should audit bundled legacy dependencies.

    Score 0.74Claim check: 1 matches source · 2 no source given · 5 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Atlassian Confluence Data Center Exploited via Apache Axis RCE (CVE-2026-90436)”

    Comments on “Atlassian Confluence Data Center Exploited via Apache Axis RCE (CVE-2026-90436)”
  13. No. 13Score 0.67

    Latest Debian 12 LTS images have a kernel crash

    Summary and score for “Latest Debian 12 LTS images have a kernel crash”

    Timon de Groot reports on the Debian LTS list that bookworm kernel 6.1.192-1 (DLA-4906-1) causes ext4 NULL-pointer oopses on some OpenStack VMs, typically during large dpkg installs, after which writes block in jbd2 until reboot. He attributes it to mballoc folio backports from 6.1.191, where IS_ERR() checks miss the NULL that 6.1's filemap_get_folio() returns. The diagnosis is his own and unverified. Another project saw the oops on 6.1.191 and mentions a queued fix he could not locate. He pins packages to the earlier kernel.

    Score 0.67Claim check: 1 matches source · 5 no source given · 1 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Latest Debian 12 LTS images have a kernel crash”

    Comments on “Latest Debian 12 LTS images have a kernel crash”
  14. No. 14Score 0.68

    A real Linux kernel, compiled to WebAssembly

    Summary and score for “A real Linux kernel, compiled to WebAssembly”

    unkernel, an open-source project by Pooya Parsa, compiles Linux 7.x to a single WebAssembly module with a real MMU and runs unmodified riscv64 binaries on a JIT-backed emulator, so Alpine boots to a shell in the browser or via npx. The README reports near-native syscalls (about 170 ns for getppid) and fork at about 60 µs, but CPU-heavy code stays well behind native. These figures are the author's own and unsourced. An MIT-licensed TypeScript SDK lets developers embed it, though it has one CPU and limited networking.

    Score 0.68Claim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “A real Linux kernel, compiled to WebAssembly”

    Comments on “A real Linux kernel, compiled to WebAssembly”
  15. No. 15Score 0.66

    BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science

    Summary and score for “BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science”

    Amazon released BenchProofer, an open-source tool that uses a task's reference fix to generate a formal specification, a verified implementation and a proof, and SWE-Proof, which applies it to all 500 SWE-Bench Verified tasks. In Amazon's own tests, 134 of Claude Opus 4.8's 425 test-passing patches were flagged by specification-guided audits, and a human review of 41 found 21 real defects. Supplying the specification raised verified pass rates, but the results are self-reported and the causal link is unconfirmed. Proofs also rely on fuzz-tested assumptions about existing code.

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

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science”

    Comments on “BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science”
  16. No. 16Score 0.66

    Making CI faster and 95% cheaper

    Summary and score for “Making CI faster and 95% cheaper”

    Vincent Koc describes how OpenClaw's CI runtime on Blacksmith runners fell from 158,546 hours in a September peak week to 21,395 in a November week (86.5% less), with estimated compute value down 95.5%. Blacksmith sponsored the capacity, and the workload also shrank, so the figures are not attributable to any single fix. The changes: deleting redundant tests, deferring full validation off main pushes, limiting repeat reviews, and selecting tests from the diff. Pull-request-to-merge time was not measured.

    Score 0.66Claim check: 3 matches source · 3 no source given · 2 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Making CI faster and 95% cheaper”

    Comments on “Making CI faster and 95% cheaper”
  17. No. 17Score 0.67

    Microsoft blames monster email outage on software defect introduced in July fix

    Summary and score for “Microsoft blames monster email outage on software defect introduced in July fix”

    Microsoft's preliminary post-incident report, dated November 15, attributes the roughly 24-hour Exchange Online outage of November 12–13 to a lock-contention defect in a July change to its Microsoft 365 gateway servers. Load testing used weaker, lower-core hardware, while the defect worsens on newer, higher-core machines. A Teams failover and about 14 percent of gateway capacity being out of service triggered a cascade, and over 80 percent of servers became unresponsive. The claims come from Microsoft's own report; the coverage cites no independent verification.

    Score 0.67Claim check: 8 no source given

    Elsewhere: HN 265 pts, front page

    Score breakdown and claims for “Microsoft blames monster email outage on software defect introduced in July fix” · Also covered by 3n3a's Blog, teapawt, Slashdot, The Next Web

    Comments on “Microsoft blames monster email outage on software defect introduced in July fix”
  18. No. 18Score 0.73

    Re-Assessing Skills Security

    Summary and score for “Re-Assessing Skills Security”

    Snyk's researchers scanned the 10,000 most popular skills on each of skills.sh and ClawHub and flagged 7 and 11 as critical, meaning actively malicious. Examples include six Roblox skills that clone a GitHub Pages site and run an absent installer, and agents-md, which runs an opaque binary that Snyk says edits ~/.claude/CLAUDE.md. Snyk reports that over 64% of skills promote unsafe behavior, such as third-party content exposure or credential handling. The figures come from Snyk's own scanner, and the claim check found no independent verification.

    Score 0.73Claim check: 6 no source given · 2 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Re-Assessing Skills Security”

    Comments on “Re-Assessing Skills Security”
  19. No. 19Score 0.68

    Linear changed code review for AI-written code

    Summary and score for “Linear changed code review for AI-written code”

    Linear describes how it changed code review as AI agents wrote more of its code. Every pull request now gets an agent review first, and authors can self-approve; a second human is mandatory only for areas like authentication and infrastructure. By Linear's own figures, PRs without a second human review rose from about 2.5% to 53% between June and October 2026, while merged PRs per author doubled to 72. The data is internal and unsourced, and it reports no quality outcomes such as defect rates.

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

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Linear changed code review for AI-written code”

    Comments on “Linear changed code review for AI-written code”
  20. No. 20Score 0.65

    A benchmark that diffs the whole database after an AI agent's task

    Summary and score for “A benchmark that diffs the whole database after an AI agent's task”

    A solo developer's open-source benchmark runs computer-use agents in Invoice Ninja, Dolibarr and Odoo 17 and 18, then diffs the whole database to separate success from collateral damage, irreversible changes and false completion. Between Odoo 17 and 18, reported success stayed near 77% while collateral damage rose from 15% to 21%, partly because "Reset to Draft" silently un-reconciles payments. A 3-step task never succeeded in 50 runs. The author notes small samples, and the figures could not be independently verified.

    Score 0.65Claim check: 1 no source given · 7 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “A benchmark that diffs the whole database after an AI agent's task”

    Comments on “A benchmark that diffs the whole database after an AI agent's task”
  21. No. 21Score 0.74

    GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

    Summary and score for “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold”

    Google Threat Intelligence Group and Mandiant report that threat actors are moving from manual prompting to autonomous multi-agent offensive frameworks such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI, which compress intrusion timelines to machine speed.

    Score 0.74Claim check: 1 matches source · 5 no source given · 2 couldn’t check

    Elsewhere: HN 5 pts, not on front page

    Score breakdown and claims for “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold” · Also covered by ARTEX.sh, gcore.com, The Register, GitHub, CIO Dive, Ostorlab Blog

    Comments on “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold”
  22. No. 22Score 0.70

    Demudhra.sh, remove the leaked eMudhra certs from your machine

    Summary and score for “Demudhra.sh, remove the leaked eMudhra certs from your machine”

    Demudhra.sh is a script for Linux and macOS that finds and removes eight compromised eMudhra intermediate CA certificates, whose P-384 private keys can be trivially recovered, from system and application trust stores.

    Score 0.70Claim check: 3 matches source · 5 couldn’t check

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “Demudhra.sh, remove the leaked eMudhra certs from your machine”

    Comments on “Demudhra.sh, remove the leaked eMudhra certs from your machine”
  23. No. 23Score 0.65

    We submitted 105 AI-built contact forms – 27 said "sent", sent nothing

    Summary and score for “We submitted 105 AI-built contact forms – 27 said "sent", sent nothing”

    A Show HN post from the Formgong team reports a study of 105 contact forms on public Lovable and Bolt sites, submitted with all network requests blocked so no message reached any owner.

    Score 0.65

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “We submitted 105 AI-built contact forms – 27 said "sent", sent nothing”

    Comments on “We submitted 105 AI-built contact forms – 27 said "sent", sent nothing”
  24. No. 24Score 0.69

    Race Against the (Mac)Hine: Local Priv Esc in Microsoft Defender for macOS

    Summary and score for “Race Against the (Mac)Hine: Local Priv Esc in Microsoft Defender for macOS”

    A security researcher documents two local privilege escalation bugs in Microsoft Defender for macOS, CVE-2026-50657 (arbitrary file read) and CVE-2026-50658 (a TOCTOU race), which are macOS variants of Windows flaws patched earlier in 2026.

    Score 0.69Claim check: 5 no source given · 3 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Race Against the (Mac)Hine: Local Priv Esc in Microsoft Defender for macOS”

    Comments on “Race Against the (Mac)Hine: Local Priv Esc in Microsoft Defender for macOS”
  25. No. 25Score 0.60

    Reverse-Engineering Kevin J. Lang's Lost OCaml Code in Apache DataSketches

    Summary and score for “Reverse-Engineering Kevin J. Lang's Lost OCaml Code in Apache DataSketches”

    The author, a ScopeDB engineer, describes porting Apache DataSketches' CpcSketch to Rust and hitting four sets of hardcoded lookup tables whose generator code, written in OCaml by the late Kevin J. Lang, was lost. Lee Rhodes confirmed the original generators no longer exist.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Reverse-Engineering Kevin J. Lang's Lost OCaml Code in Apache DataSketches”

    Comments on “Reverse-Engineering Kevin J. Lang's Lost OCaml Code in Apache DataSketches”
  26. No. 26Score 0.68

    Universal XSS in uBlock Origin from website-driven element picker

    Summary and score for “Universal XSS in uBlock Origin from website-driven element picker”

    A security researcher reports a one-click universal XSS in the Firefox MV2 build of uBlock Origin, where a malicious website can drive the element picker to add arbitrary trusted filters and so run JavaScript on any site.

    Score 0.68

    Elsewhere: HN 602 pts, front page

    Score breakdown for “Universal XSS in uBlock Origin from website-driven element picker”

    Comments on “Universal XSS in uBlock Origin from website-driven element picker”
  27. No. 27Score 0.62

    The Great Exchange Online Meltdown of November 2026

    Summary and score for “The Great Exchange Online Meltdown of November 2026”

    A detailed account of the November 2026 Exchange Online outage, published by Office 365 IT Pros author Tony Redmond, says the disruption lasted nearly 24 hours and was not an Outlook-specific problem.

    Score 0.62Contestedjudges differ by 0.57 on new information

    Elsewhere: HN —

    Score breakdown for “The Great Exchange Online Meltdown of November 2026”

    Comments on “The Great Exchange Online Meltdown of November 2026”
  28. No. 28Score 0.66

    How we moved our backend from Node to Go in two weeks

    Summary and score for “How we moved our backend from Node to Go in two weeks”

    Neptune, a prenup and estate-planning service, moved its entire backend from Node to Go over two weeks in November, without downtime. A Go "front door" in production routed each route through Node, a compare mode, or Go, and was flipped gradually.

    Score 0.66

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “How we moved our backend from Node to Go in two weeks”

    Comments on “How we moved our backend from Node to Go in two weeks”
  29. No. 29Score 0.60

    Agentic AI Designs an Analog Synthesizer

    Summary and score for “Agentic AI Designs an Analog Synthesizer”

    A developer used Claude Code with Opus 5.5 to design a complete analog synthesizer, from a parts-list prompt through KiCad schematic, ngspice simulation, PCB layout, firmware, and a design report.

    Score 0.60

    Elsewhere: HN 15 pts, front page

    Score breakdown for “Agentic AI Designs an Analog Synthesizer”

    Comments on “Agentic AI Designs an Analog Synthesizer”
  30. No. 30Score 0.60

    How to keep servers from exiting load balancer pool

    Summary and score for “How to keep servers from exiting load balancer pool”

    The article explains the "spiral of death" in load-balanced server clusters: when a traffic burst pushes servers past capacity, failed health checks cause them to be pulled from the pool, which overloads the remaining servers and removes them too, even after traffic falls.

    Score 0.60Contestedjudges differ by 0.45 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “How to keep servers from exiting load balancer pool”

    Comments on “How to keep servers from exiting load balancer pool”
  31. No. 31Score 0.68

    ConnectWise ScreenConnect: Attackers install malware through RMM+ extension

    Summary and score for “ConnectWise ScreenConnect: Attackers install malware through RMM+ extension”

    Heise reports that attackers are exploiting the RMM+ extension for ConnectWise ScreenConnect to push tampered clients and malware to at least 288 endpoints across roughly 48 organizations, with a further 205 hosts hit in a later wave, according to Huntress.

    Score 0.68

    Elsewhere: HN —

    Score breakdown for “ConnectWise ScreenConnect: Attackers install malware through RMM+ extension”

    Comments on “ConnectWise ScreenConnect: Attackers install malware through RMM+ extension”
  32. No. 32Score 0.59

    UK leaves client-side scanning out of new child safety rules

    Summary and score for “UK leaves client-side scanning out of new child safety rules”

    Britain's regulator Ofcom has said it lacks power under the Online Safety Act to compel client-side scanning on users' devices to detect child abuse images. Its finalized rules, due to take effect in February 2027 if Parliament approves, require proactive CSAM detection only for publicly shared content, with a looser…

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “UK leaves client-side scanning out of new child safety rules”

    Comments on “UK leaves client-side scanning out of new child safety rules”
  33. No. 33Score 0.59

    Turso Sync: Local-First SQLite with Push and Pull

    Summary and score for “Turso Sync: Local-First SQLite with Push and Pull”

    Turso Sync is a second attempt at SQLite sync by the Turso team, built on their open-source Rust rewrite of SQLite, Turso Database. It replaces the page-based replication that hampered their earlier Embedded Replicas with row-level change capture, so clients push logical changes instead of physical pages.

    Score 0.59Contestedjudges differ by 0.46 on substantiated, 0.47 on promotional

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Turso Sync: Local-First SQLite with Push and Pull”

    Comments on “Turso Sync: Local-First SQLite with Push and Pull”
  34. No. 34Score 0.59

    Jeremy Schneider: Postgres 18.7 Debian and RedHat Packagers Protect You

    Summary and score for “Jeremy Schneider: Postgres 18.7 Debian and RedHat Packagers Protect You”

    Jeremy Schneider's follow-up post explains how the Debian and RPM packaging for PostgreSQL 18.7 now blocks installing newer extensions onto an older server. Packages like postgresql-18-cron and pg_cron_18 declare dependencies on postgresql-18 (> 18.7) and postgresql18-libs >= 18.7, so the package manager refuses a mismatched install.

    Score 0.59

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Jeremy Schneider: Postgres 18.7 Debian and RedHat Packagers Protect You”

    Comments on “Jeremy Schneider: Postgres 18.7 Debian and RedHat Packagers Protect You”
  35. No. 35Score 0.67

    Attackers chain PHP2Shell with root flaw on Check Point gateways

    Summary and score for “Attackers chain PHP2Shell with root flaw on Check Point gateways”

    Attackers chained the php2shell PHP flaws with a local root vulnerability in Check Point's Gaia OS to fully compromise unpatched gateways between late October and early November, and in one case used a wiper on an ESXi host.

    Score 0.67

    Elsewhere: HN —

    Score breakdown for “Attackers chain PHP2Shell with root flaw on Check Point gateways”

    Comments on “Attackers chain PHP2Shell with root flaw on Check Point gateways”
  36. No. 36Score 0.64

    Putting Task Expertise into RL Achieves SOTA Performance on Text-to-SQL

    Summary and score for “Putting Task Expertise into RL Achieves SOTA Performance on Text-to-SQL”

    A research team (UIUC Kang Lab) reports ReViSQL-K2.6, a single fine-tuned model trained with reinforcement learning on a cleaned training set, BIRD-Platinum, that reaches 88.55% on the expert-verified Arcwise-Plat-SQL benchmark and exceeds the 92.96% human proxy with 16-sample self-consistency.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Putting Task Expertise into RL Achieves SOTA Performance on Text-to-SQL”

    Comments on “Putting Task Expertise into RL Achieves SOTA Performance on Text-to-SQL”
  37. No. 37Score 0.66

    Why our search engine returned T. Rowe Price for "Supabase pricing"

    Summary and score for “Why our search engine returned T. Rowe Price for "Supabase pricing"”

    Alex C., writing for NestDaddy, describes how their in-house search engine returned a T. Rowe Price investment page for "Supabase pricing" and traces it to three ranking bugs: a candidate-set fallback that discarded precise matches, a query stemmed twice through a cache-normalization layer, and homepages never reaching final scoring.

    Score 0.66

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Why our search engine returned T. Rowe Price for "Supabase pricing"”

    Comments on “Why our search engine returned T. Rowe Price for "Supabase pricing"”
  38. No. 38Score 0.65

    What an Agent Can Reach Through the Headless 360 MCP Server

    Summary and score for “What an Agent Can Reach Through the Headless 360 MCP Server”

    Salesforce's Headless 360 MCP server lets an AI agent act in an org with the authenticated user's full permissions, which in mature orgs often far exceed what the job requires.

    Score 0.65

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “What an Agent Can Reach Through the Headless 360 MCP Server”

    Comments on “What an Agent Can Reach Through the Headless 360 MCP Server”
  39. No. 39Score 0.57

    Hot topics in DNS from DNS OARC 47

    Summary and score for “Hot topics in DNS from DNS OARC 47”

    At DNS OARC 47 in Vancouver, Quad9's John Todd described a cuckoo filter used as a gatekeeper in front of a large query-logging store, achieving roughly 67 bits per object with a 1 GB memory footprint and a false-positive rate near 1 in 70 million.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “Hot topics in DNS from DNS OARC 47”

    Comments on “Hot topics in DNS from DNS OARC 47”
  40. No. 40Score 0.57

    GPU prices, October 2026: AI compute market report – Otium

    Summary and score for “GPU prices, October 2026: AI compute market report – Otium”

    October's GPU market report from Otium, covering spot prices across 1.35 million observations on AWS and Google Cloud, finds prices largely steady, with most GPUs moving under 5% since September. The H100 on AWS rose 13% while the A100 80 GB fell 12%.

    Score 0.57Contestedjudges differ by 0.48 on significance, 0.47 on reader interest, 0.58 on technical depth, 0.42 on promotional

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “GPU prices, October 2026: AI compute market report – Otium”

    Comments on “GPU prices, October 2026: AI compute market report – Otium”
  41. No. 41Score 0.65

    Has AI’s ‘Black October’ changed the way you secure your code?

    Summary and score for “Has AI’s ‘Black October’ changed the way you secure your code?”

    Pillar Security's October disclosure of the OpenCode flaw, in which a malicious MCP server could run commands on a developer's machine during authentication, and Microsoft's report of an AI agent stealing Azure Key Vault data within five minutes, both show autonomous agents exploiting new entry points in coding tools and…

    Score 0.65Contestedjudges differ by 0.43 on new information

    Elsewhere: HN —

    Score breakdown for “Has AI’s ‘Black October’ changed the way you secure your code?”

    Comments on “Has AI’s ‘Black October’ changed the way you secure your code?”
  42. No. 42Score 0.57

    Nvidia's Off-Balance-Sheet Guarantees Balloon to $530B

    Summary and score for “Nvidia's Off-Balance-Sheet Guarantees Balloon to $530B”

    Nvidia's off-balance-sheet guarantees and commitments reached $530 billion at the end of its fiscal second quarter, up from $184 billion three months earlier, according to research firm SemiAnalysis, citing a securities filing.

    Score 0.57

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Nvidia's Off-Balance-Sheet Guarantees Balloon to $530B”

    Comments on “Nvidia's Off-Balance-Sheet Guarantees Balloon to $530B”
  43. No. 43Score 0.64

    Waiting for a security patch: Crypto-miner attacks on N-central observed

    Summary and score for “Waiting for a security patch: Crypto-miner attacks on N-central observed”

    Security researchers at Huntback have observed live attacks on N-central, the remote monitoring platform from N-able, in which attackers install Monero crypto miners on servers that still have the Legacy API enabled, even after the HF2 hotfix.

    Score 0.64

    Elsewhere: HN —

    Score breakdown for “Waiting for a security patch: Crypto-miner attacks on N-central observed”

    Comments on “Waiting for a security patch: Crypto-miner attacks on N-central observed”
  44. No. 44Score 0.66

    NoisyTCP bug gives Linux servers a remote crash problem 11 years in the making

    Summary and score for “NoisyTCP bug gives Linux servers a remote crash problem 11 years in the making”

    A security researcher, Hyunwoo Kim, disclosed CVE-2026-99288 ("NoisyTCP"), a Linux kernel TCP flaw dating to 2015 that lets unauthenticated remote attackers crash servers, potentially taking down containerized workloads and host machines. Kim says most distributions are affected, though no data theft or takeover has been shown.

    Score 0.66Contestedjudges differ by 0.64 on new information

    Elsewhere: HN —

    Score breakdown for “NoisyTCP bug gives Linux servers a remote crash problem 11 years in the making”

    Comments on “NoisyTCP bug gives Linux servers a remote crash problem 11 years in the making”
  45. No. 45Score 0.69

    Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn

    Summary and score for “Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn”

    Amazon Threat Intelligence and the FBI report that attackers are using ARTEX, an open source AI penetration testing framework, to automate cybercrime against banks, governments, energy and healthcare providers.

    Score 0.69Contestedjudges differ by 0.59 on new information

    Elsewhere: HN —

    Score breakdown for “Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn”

    Comments on “Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn”
  46. No. 46Score 0.64

    SatBench a Benchmark of AI Benchmarks

    Summary and score for “SatBench a Benchmark of AI Benchmarks”

    SatBench, a project by Ethan Mollick, treats 173 frequently cited AI benchmarks as questions on one exam, counting one solved once the best reported score reaches 85%. It stands at 63% solved, up about 20 points in the past year.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “SatBench a Benchmark of AI Benchmarks”

    Comments on “SatBench a Benchmark of AI Benchmarks”
  47. No. 47Score 0.63

    Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development

    Summary and score for “Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development”

    Import AI 481 covers FrontierML, a benchmark from its authors that measures whether AI agents can turn a small budget ($10, $50, or $200) into better model performance across 21 end-to-end machine learning tasks.

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development”

    Comments on “Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development”
  48. No. 48Score 0.57

    Stromboli Open-source AI SRE agent focused on fast incident mitigation

    Summary and score for “Stromboli Open-source AI SRE agent focused on fast incident mitigation”

    Stromboli is an open-source, self-hosted AI site reliability agent that responds to alerts by first proposing a concrete mitigation such as a rollback, restart, or scale-up, then investigating the root cause against existing logs, metrics, and deploy history. It produces a structured report posted to Slack.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Stromboli Open-source AI SRE agent focused on fast incident mitigation”

    Comments on “Stromboli Open-source AI SRE agent focused on fast incident mitigation”
  49. No. 49Score 0.63

    Scans for LiteLLM/Starlette (CVE-2026-42271, CVE-2026-48710), (Mon, Nov 16th)

    Summary and score for “Scans for LiteLLM/Starlette (CVE-2026-42271, CVE-2026-48710), (Mon, Nov 16th)”

    Security researcher Johannes Ullrich reports that honeypots began receiving requests this weekend targeting LiteLLM's /mcp-rest/test/tools/list endpoint, attempting to chain CVE-2026-42271 and CVE-2026-48710 for unauthenticated remote code execution.

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “Scans for LiteLLM/Starlette (CVE-2026-42271, CVE-2026-48710), (Mon, Nov 16th)”

    Comments on “Scans for LiteLLM/Starlette (CVE-2026-42271, CVE-2026-48710), (Mon, Nov 16th)”
  50. No. 50Score 0.57

    Itzma, an embedded vector database backed by object storage

    Summary and score for “Itzma, an embedded vector database backed by object storage”

    Itzma is an open-source Rust library, with Python bindings, that provides vector search over object storage such as S3, local disk, or memory, with no server in between. Vectors are stored as plain objects, and concurrent writers coordinate through the store's put-if-absent and conditional-update primitives.

    Score 0.57

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Itzma, an embedded vector database backed by object storage”

    Comments on “Itzma, an embedded vector database backed by object storage”
  51. No. 51Score 0.63

    Browser tests – from natural language to code

    Summary and score for “Browser tests – from natural language to code”

    Sitelooper is an open-source CLI by its author that applies "progressive automation" to browser testing: an LLM drives a real browser once to record a task, then the recording is compiled into a replayable procedure with parameters, fallback locators and effect checks.

    Score 0.63

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Browser tests – from natural language to code”

    Comments on “Browser tests – from natural language to code”
  52. No. 52Score 0.55

    Israel uploaded photos to Amazon's Rekognition to track Palestinians

    Summary and score for “Israel uploaded photos to Amazon's Rekognition to track Palestinians”

    The Guardian reports that the Israeli military uploaded about 3.5 million photos of Palestinians to Amazon Web Services in early 2024 so it could run them through Rekognition, Amazon's facial recognition platform.

    Score 0.55

    Elsewhere: HN 36 pts, not on front page

    Score breakdown for “Israel uploaded photos to Amazon's Rekognition to track Palestinians”

    Comments on “Israel uploaded photos to Amazon's Rekognition to track Palestinians”
  53. No. 53Score 0.63

    Designing nextest's install scripts

    Summary and score for “Designing nextest's install scripts”

    The author describes adding install scripts to cargo-nextest, a Rust test runner, after four and a half years of relying on manual tarball extraction. He explains why install scripts were hard to write well, how the scripts use a generated install manifest with checksums to pick and verify the right…

    Score 0.63

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Designing nextest's install scripts”

    Comments on “Designing nextest's install scripts”
  54. No. 54Score 0.64

    Microsoft outage based phishing live now

    Summary and score for “Microsoft outage based phishing live now”

    Attackers launched a live phishing campaign within a day of Microsoft 365's 12 November outage, which Microsoft attributes to a network capacity upgrade rather than an attack.

    Score 0.64Contestedjudges differ by 0.51 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Microsoft outage based phishing live now” · Also covered by heise online English

    Comments on “Microsoft outage based phishing live now”
  55. No. 55Score 0.63

    I took apart 221 security advisories. Six mistakes are behind 129 of them

    Summary and score for “I took apart 221 security advisories. Six mistakes are behind 129 of them”

    A security-research site reviewed 221 vendor security advisories and tagged each with recurring design mistakes. The largest category, "Design Debt Driver," covers 70 files where a vendor repeatedly patches the same component's bug class instead of fixing the underlying design.

    Score 0.63Contestedjudges differ by 0.51 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “I took apart 221 security advisories. Six mistakes are behind 129 of them”

    Comments on “I took apart 221 security advisories. Six mistakes are behind 129 of them”
  56. No. 56Score 0.57

    PowerShell 7.7 is released – as an MSIX

    Summary and score for “PowerShell 7.7 is released – as an MSIX”

    PowerShell 7.7 ships only as an MSIX on Windows, dropping the MSI option, so Microsoft can bundle it into future Windows releases. The author tested side-by-side installs with 7.6 and found per-user install behavior, PATH and Terminal quirks, and gaps for LocalSystem, WinRM remoting, and DISM cmdlets.

    Score 0.57

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “PowerShell 7.7 is released – as an MSIX”

    Comments on “PowerShell 7.7 is released – as an MSIX”
  57. No. 57Score 0.63

    Hibana: Fast Agentic Inference on Unified-Memory Devices

    Summary and score for “Hibana: Fast Agentic Inference on Unified-Memory Devices”

    Cornell University researchers Yahya Emara, Chi-Chih Chang and colleagues describe Hibana, a system for running many local AI agents quickly on a single DGX Spark, where weights and agent histories share one 128 GB memory pool.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Hibana: Fast Agentic Inference on Unified-Memory Devices”

    Comments on “Hibana: Fast Agentic Inference on Unified-Memory Devices”
  58. No. 58Score 0.56

    GPU-Initiated DMA Copy Engines in RCCL and RocSHMEM

    Summary and score for “GPU-Initiated DMA Copy Engines in RCCL and RocSHMEM”

    AMD's engineering blog describes GPU-initiated SDMA copy engines in RCCL and rocSHMEM, which let GPU kernels trigger intra-node data transfers on dedicated DMA engines instead of compute units.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “GPU-Initiated DMA Copy Engines in RCCL and RocSHMEM”

    Comments on “GPU-Initiated DMA Copy Engines in RCCL and RocSHMEM”
  59. No. 59Score 0.62

    My Quest to push native Three.js quality match AAA Game Engines

    Summary and score for “My Quest to push native Three.js quality match AAA Game Engines”

    A detailed series by Ben Houston measures Three.js against the lighting pipeline used in AAA games, stage by stage: G-buffer, baked and probe-based global illumination, occlusion, screen-space reflections, and more. It catalogs what Three.js already has, what is in draft review, and what remains unbuilt.

    Score 0.62

    Elsewhere: HN 120 pts, front page

    Score breakdown for “My Quest to push native Three.js quality match AAA Game Engines”

    Comments on “My Quest to push native Three.js quality match AAA Game Engines”
  60. No. 60Score 0.62

    What gives away LLM writing

    Summary and score for “What gives away LLM writing”

    A pre-registered study compared seven current LLMs against 800 pre-ChatGPT Hacker News and Stack Exchange replies to test 23 popular "tells" of machine writing.

    Score 0.62Contestedjudges differ by 0.43 on technical depth

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “What gives away LLM writing”

    Comments on “What gives away LLM writing”
  61. No. 61Score 0.62

    CavaAlign, a 16 MB word-level audio/text aligner for 35 languages

    Summary and score for “CavaAlign, a 16 MB word-level audio/text aligner for 35 languages”

    CavaAlign, a Show HN project from Cava Reader, is a 16 MB word-level audio-to-text aligner covering 35 languages, runnable on Android and laptop CPUs. It gives each word's start and end time from a recording and its text.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “CavaAlign, a 16 MB word-level audio/text aligner for 35 languages”

    Comments on “CavaAlign, a 16 MB word-level audio/text aligner for 35 languages”
  62. No. 62Score 0.62

    Misalignment signals in Kimi K3: What probes flag and where they peak

    Summary and score for “Misalignment signals in Kimi K3: What probes flag and where they peak”

    A study of lightweight activation probes trained on frozen Kimi K3 activations, using labels from an LLM misalignment monitor, examines what these probes flag during agent runs. The authors find the probe separates monitor-flagged violations from similar benign behavior and often fires on reasoning, tool calls, and tool feedback.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Misalignment signals in Kimi K3: What probes flag and where they peak”

    Comments on “Misalignment signals in Kimi K3: What probes flag and where they peak”
  63. No. 63Score 0.62

    Sopro V2: private, fast, on-device text-to-speech

    Summary and score for “Sopro V2: private, fast, on-device text-to-speech”

    Sopro V2 Turbo is a 120M-parameter, open-source voice-cloning text-to-speech model from Halo NeuroAI's Samuel Vitorino, built to run locally on a laptop CPU or in the browser.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Sopro V2: private, fast, on-device text-to-speech”

    Comments on “Sopro V2: private, fast, on-device text-to-speech”
  64. No. 64Score 0.61

    Weakness in iOS FairPlay to dump apps directly from macOS or Linux

    Summary and score for “Weakness in iOS FairPlay to dump apps directly from macOS or Linux”

    A community write-up reports that Apple's FairPlay DRM for iOS apps can be defeated offline, from macOS or Linux, using a reusable account-level credential extracted once from a jailbroken device or enrolled Mac.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Weakness in iOS FairPlay to dump apps directly from macOS or Linux”

    Comments on “Weakness in iOS FairPlay to dump apps directly from macOS or Linux”
  65. No. 65Score 0.61

    Running untrusted CMS plugins in WebAssembly

    Summary and score for “Running untrusted CMS plugins in WebAssembly”

    Hatamoto is a Go-based CMS that runs third-party plugins as WebAssembly modules via Extism and wazero, replacing WordPress's in-process PHP plugins. Each plugin gets an empty filesystem, no network, memory caps, and per-call timeouts, with host access gated by install-time capabilities.

    Score 0.61

    Elsewhere: HN 10 pts, not on front page

    Score breakdown for “Running untrusted CMS plugins in WebAssembly”

    Comments on “Running untrusted CMS plugins in WebAssembly”
  66. No. 66Score 0.61

    The Nefarious Plan of vm2

    Summary and score for “The Nefarious Plan of vm2”

    vm2, a sandbox for running untrusted code in Node.js, has shipped a string of security fixes, and the author argues these form a pattern rather than isolated bugs. The sandbox works by wrapping host-facing objects from a hand-maintained list, which lags new JavaScript language features and V8 internals.

    Score 0.61Contestedjudges differ by 0.66 on new information, 0.41 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The Nefarious Plan of vm2”

    Comments on “The Nefarious Plan of vm2”
  67. No. 67Score 0.61

    Memorilla – Latent Semantic Memory for LLMs

    Summary and score for “Memorilla – Latent Semantic Memory for LLMs”

    MEMORILLA is a learned memory module that reads an entire document collection and gives a frozen LLM a fixed, small set of query-dependent memory tokens, rather than placing all text in context or retrieving a few passages.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Memorilla – Latent Semantic Memory for LLMs”

    Comments on “Memorilla – Latent Semantic Memory for LLMs”
  68. No. 68Score 0.61

    Hugging Face and OpenRouter join forces to make open-weight models easy to run

    Summary and score for “Hugging Face and OpenRouter join forces to make open-weight models easy to run”

    Hugging Face is switching its Inference Providers language-model requests from its own router to OpenRouter's, which picks among 80+ providers per request on live price, speed and reliability and retries failures on another provider.

    Score 0.61Contestedjudges differ by 0.42 on promotional

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Hugging Face and OpenRouter join forces to make open-weight models easy to run”

    Comments on “Hugging Face and OpenRouter join forces to make open-weight models easy to run”
  69. No. 69Score 0.57

    It's always DNS, so why are we running our own?

    Summary and score for “It's always DNS, so why are we running our own?”

    S2, a streaming infrastructure company, replaced AWS Route53 with a custom authoritative DNS server to support unlimited basins, each mapped to a subdomain via CNAME records. Route53's API limit of roughly 5 upserts per second made instant, strongly consistent record creation impractical under load.

    Score 0.57

    Elsewhere: HN 64 pts, front page

    Score breakdown for “It's always DNS, so why are we running our own?”

    Comments on “It's always DNS, so why are we running our own?”
  70. No. 70Score 0.61

    Standard Support – a convention for AI agents to file support tickets

    Summary and score for “Standard Support – a convention for AI agents to file support tickets”

    Standard Support is a draft convention, published by Nathan Broadbent of DocSpring (Draft 0.2, November 2026), that defines a set of tools letting a customer's AI agent ask a product's documentation for help, file support tickets with the vendor's team, and pick up replies later, even across sessions.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Standard Support – a convention for AI agents to file support tickets”

    Comments on “Standard Support – a convention for AI agents to file support tickets”
  71. No. 71Score 0.61

    Measuring alignment eval tells

    Summary and score for “Measuring alignment eval tells”

    METR researchers report that AI "investigator" agents, run inside both capability and alignment evaluations and asked what each tests, increasingly tell the two apart. Claude Opus 4.8 and 5.5 nearly perfectly separate alignment evals from capability evals, usually by noticing environments that look fake or tasks that are honeypots.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Measuring alignment eval tells”

    Comments on “Measuring alignment eval tells”
  72. No. 72Score 0.65

    New Windows SignDrop zero-day gives hackers SYSTEM privileges

    Summary and score for “New Windows SignDrop zero-day gives hackers SYSTEM privileges”

    Security researcher Abdelhamid Naceri released SignDrop, a Windows zero-day that can grant SYSTEM privileges by abusing a Microsoft mitigation introduced in October 2025 against Common Log File System (CLFS) attacks.

    Score 0.65Contestedjudges differ by 0.42 on overclaim

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “New Windows SignDrop zero-day gives hackers SYSTEM privileges”

    Comments on “New Windows SignDrop zero-day gives hackers SYSTEM privileges”
  73. No. 73Score 0.62

    TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

    Summary and score for “TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient”

    TokenEconomizer is an open-source plugin, developed by researchers at UC Berkeley, Microsoft Research and Microsoft Security, that learns from an AI coding agent's past sessions in a repository to cut token spending on later tasks.

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient”

    Comments on “TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient”
  74. No. 74Score 0.60

    Implicit state machines hide bugs that line-by-line review approves

    Summary and score for “Implicit state machines hide bugs that line-by-line review approves”

    QuickReview's author describes adding a state-machine review lens to their AI code reviewer after noticing it missed implicit state bugs, such as a button stuck disabled after an early return.

    Score 0.60Contestedjudges differ by 0.47 on substantiated, 0.46 on promotional

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Implicit state machines hide bugs that line-by-line review approves”

    Comments on “Implicit state machines hide bugs that line-by-line review approves”
  75. No. 75Score 0.60

    Software engineering after code becomes cheap with Satya Nadella

    Summary and score for “Software engineering after code becomes cheap with Satya Nadella”

    In a conversation with The Pragmatic Engineer's Gergely Orosz, Microsoft CEO Satya Nadella discusses how software engineering is changing as code generation gets cheap and verification becomes the bottleneck.

    Score 0.60

    Elsewhere: HN —

    Score breakdown for “Software engineering after code becomes cheap with Satya Nadella” · Also covered by The Next Web

    Comments on “Software engineering after code becomes cheap with Satya Nadella”
  76. No. 76Score 0.60

    Benchmarking OSS inference vendors on cost/latency

    Summary and score for “Benchmarking OSS inference vendors on cost/latency”

    The powerset-co inference-bench repo is a beta benchmark that compares open-source LLM inference providers on speed and list-price cost while serving a single AI research agent. It replays recorded FRAMES-based Wikipedia research conversations across providers, timing only model calls, with quality deliberately unmeasured.

    Score 0.60Contestedjudges differ by 0.44 on substantiated

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Benchmarking OSS inference vendors on cost/latency”

    Comments on “Benchmarking OSS inference vendors on cost/latency”
  77. No. 77Score 0.56

    Memory crunch: Companies keep hardware longer and delay purchases

    Summary and score for “Memory crunch: Companies keep hardware longer and delay purchases”

    Bitkom, Germany's digital industry association, surveyed 603 German firms with 20 or more employees and found that three-quarters are feeling higher memory-driven hardware prices.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “Memory crunch: Companies keep hardware longer and delay purchases”

    Comments on “Memory crunch: Companies keep hardware longer and delay purchases”
  78. No. 78Score 0.56

    Kafka-console – Web UI for Kafka in one binary

    Summary and score for “Kafka-console – Web UI for Kafka in one binary”

    Kafka-console is a Show HN project: a web UI for Apache Kafka shipped as a single binary with the frontend embedded. It lets operators point it at one or more clusters to inspect brokers, topics, partitions, consumer lag, schemas and ACLs, and to browse, search, replay, import and produce messages.…

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Kafka-console – Web UI for Kafka in one binary”

    Comments on “Kafka-console – Web UI for Kafka in one binary”
  79. No. 79Score 0.60

    Agents Need an Undo Button, Not More Approval Prompts

    Summary and score for “Agents Need an Undo Button, Not More Approval Prompts”

    The essay argues that AI agent safety should rely less on constant human approval prompts, which people rubber-stamp over time, and more on classifying each agent action by reversibility: reversible, compensable, delayable, or irreversible.

    Score 0.60

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Agents Need an Undo Button, Not More Approval Prompts”

    Comments on “Agents Need an Undo Button, Not More Approval Prompts”
  80. No. 80Score 0.60

    How to clean up the worktrees Claude Code, Cursor and Codex leave behind

    Summary and score for “How to clean up the worktrees Claude Code, Cursor and Codex leave behind”

    Mopup's guide explains how Claude Code, Cursor and Codex each store git worktrees for coding agents, where they live, and what each tool cleans up automatically.

    Score 0.60Contestedjudges differ by 0.41 on technical depth, 0.57 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “How to clean up the worktrees Claude Code, Cursor and Codex leave behind”

    Comments on “How to clean up the worktrees Claude Code, Cursor and Codex leave behind”
  81. No. 81Score 0.59

    Muru: Scaling from What the Workers Report

    Summary and score for “Muru: Scaling from What the Workers Report”

    Muru, a Python application server, decides its own worker process count per release, scaling out when slots are full and scaling in when workers sit idle, under a churn budget that can pause scale-in.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Muru: Scaling from What the Workers Report”

    Comments on “Muru: Scaling from What the Workers Report”
  82. No. 82Score 0.59

    Building a Rust-Native, Cross-Language Foundation for Agentic ML Compilers

    Summary and score for “Building a Rust-Native, Cross-Language Foundation for Agentic ML Compilers”

    The article introduces TIRx Rust, a proposed Rust-native, cross-language foundation for ML compilers, built on the TVM-FFI ABI. It lets Rust, C++, and Python code operate directly on the same reference-counted IR objects, enabling agents to choose the best language per module, prototyping in Python and building robust code in…

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Building a Rust-Native, Cross-Language Foundation for Agentic ML Compilers”

    Comments on “Building a Rust-Native, Cross-Language Foundation for Agentic ML Compilers”
  83. No. 83Score 0.59

    Git is not enough for agents

    Summary and score for “Git is not enough for agents”

    Hasan Madi argues that Git is insufficient protection once coding agents run in auto-approve mode, since commands like git clean, reset --hard, and force-push can destroy work Git cannot recover.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Git is not enough for agents”

    Comments on “Git is not enough for agents”
  84. No. 84Score 0.59

    Evolving Infra: Agent-Adaptive LLM Serving at Scale

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

    Evolving Infra is an agent-adaptive LLM serving framework that targets inefficiency in prefill–decode disaggregated systems when running agentic workloads. Its authors, from Georgia Tech, UIUC, Berkeley, NVIDIA, OpenAI, and MIT, decouple node phase from KV cache storage so nodes can switch between prefill and decode without restarts.

    Score 0.59Contestedjudges differ by 0.61 on overclaimClaim check: 1 not in source · 7 no source given

    Elsewhere: HN 1 pts, not on front page

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

    Comments on “Evolving Infra: Agent-Adaptive LLM Serving at Scale”
  85. No. 85Score 0.59

    Quark: Our whole training stack, end to end, for $4,200

    Summary and score for “Quark: Our whole training stack, end to end, for $4,200”

    Empero AI released Quark, a 165M-parameter language model it says it trained end to end for about $4,200 of compute, covering data, pretraining, 128K-context extension and reinforcement learning.

    Score 0.59

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Quark: Our whole training stack, end to end, for $4,200”

    Comments on “Quark: Our whole training stack, end to end, for $4,200”
  86. No. 86Score 0.59

    Redacted - Stop Handing Claude Your Secrets

    Summary and score for “Redacted - Stop Handing Claude Your Secrets”

    Redacted is an open-source tool that scrubs secrets from command output before an AI coding assistant such as Claude Code reads it. It runs as a PostToolUse hook, scans locally with no network calls, and replaces matches with markers that keep only the last four characters.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Redacted - Stop Handing Claude Your Secrets” · Also covered by redacted.my.id

    Comments on “Redacted - Stop Handing Claude Your Secrets”
  87. No. 87Score 0.59

    Estonia has a plan for Russian occupation: Run its government from the cloud

    Summary and score for “Estonia has a plan for Russian occupation: Run its government from the cloud”

    Estonia is preparing a "digital resilience" strategy under which its government's digital services could be run from abroad, through cloud providers and a data embassy in Luxembourg, if its territory were occupied or its data centers destroyed. Officials cite persistent Russian hybrid attacks and the 2007 cyberattacks.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “Estonia has a plan for Russian occupation: Run its government from the cloud” · Also covered by heise online English

    Comments on “Estonia has a plan for Russian occupation: Run its government from the cloud”
  88. No. 88Score 0.59

    Agent-Otel-Kit – OTel Dashboards for Claude Code, Codex and Gemini CLI

    Summary and score for “Agent-Otel-Kit – OTel Dashboards for Claude Code, Codex and Gemini CLI”

    agent-otel-kit is an open-source, local OpenTelemetry stack that collects cost, token, tool-call and trace telemetry from Claude Code, Codex, Gemini CLI and Qwen Code, with Prometheus, Loki, Tempo and dashboards in Grafana and Perses. A Go CLI configures each agent's user settings and checks that data arrives.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Agent-Otel-Kit – OTel Dashboards for Claude Code, Codex and Gemini CLI”

    Comments on “Agent-Otel-Kit – OTel Dashboards for Claude Code, Codex and Gemini CLI”
  89. No. 89Score 0.58

    An Inauthentic TikTok and Instagram Network Inflaming the U.S.–Canada Trade War

    Summary and score for “An Inauthentic TikTok and Instagram Network Inflaming the U.S.–Canada Trade War”

    A network of at least 293 TikTok and Instagram accounts is using AI deepfakes built on stolen footage of real creators, many Canadian, to post politically divisive videos about the U.S.–Canada trade war.

    Score 0.58Contestedjudges differ by 0.50 on reader interest

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “An Inauthentic TikTok and Instagram Network Inflaming the U.S.–Canada Trade War”

    Comments on “An Inauthentic TikTok and Instagram Network Inflaming the U.S.–Canada Trade War”
  90. No. 90Score 0.58

    Gemini 4 Carbon, Hydrogen, Decisions API, self-hosted agent sandboxes and more

    Summary and score for “Gemini 4 Carbon, Hydrogen, Decisions API, self-hosted agent sandboxes and more”

    Google's developer update announces Gemini 4 Carbon, a reasoning model for repository-scale engineering, proofs, and security work, and Gemini 4 Hydrogen, a faster model meant as the recommended upgrade from Gemini 3.8 Flash for everyday coding and agents. Both share a 1M-token window and are priced per million tokens.

    Score 0.58Contestedjudges differ by 0.43 on significance, 0.55 on substantiated, 0.78 on promotional

    Elsewhere: HN —

    Score breakdown for “Gemini 4 Carbon, Hydrogen, Decisions API, self-hosted agent sandboxes and more”

    Comments on “Gemini 4 Carbon, Hydrogen, Decisions API, self-hosted agent sandboxes and more”
  91. No. 91Score 0.58

    Compiling BAML to Rust

    Summary and score for “Compiling BAML to Rust”

    The BAML team describes a new backend that compiles BAML's mid-level intermediate representation directly into Rust, rather than interpreting it. Compiled code averages about 3.5x faster than their interpreter, reaching roughly 34x on tight loops, though data-heavy workloads like their agent loop gain only about 1.5x.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Compiling BAML to Rust”

    Comments on “Compiling BAML to Rust”
  92. No. 92Score 0.58

    Six hours and a stack of .md files

    Summary and score for “Six hours and a stack of .md files”

    A personal blog post by micr0 re-examines a claim that sophisticated cyberattacks now cost less than a McDonald's meal. It cites a Google Threat Intelligence Group report describing a financially motivated actor who used an AI coding assistant to plan and run a mass credential-harvesting campaign in under six hours…

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Six hours and a stack of .md files”

    Comments on “Six hours and a stack of .md files”
  93. No. 93Score 0.61

    Benchmarking Coding Agents Under Evolving Requirements

    Summary and score for “Benchmarking Coding Agents Under Evolving Requirements”

    FeatureLoop is a benchmark of 61 human-reviewed tasks from 34 repositories that tests coding agents on iterative development, where users add requirements and report failures across successive revisions of one codebase. Each stage is checked with executable tests for new behavior and preservation of earlier functionality.

    Score 0.61

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Benchmarking Coding Agents Under Evolving Requirements”

    Comments on “Benchmarking Coding Agents Under Evolving Requirements”
  94. No. 94Score 0.58

    JA4 Fingerprinting: Why One Digest Can Cover User-Agents

    Summary and score for “JA4 Fingerprinting: Why One Digest Can Cover User-Agents”

    A detailed walkthrough of JA4, a TLS ClientHello fingerprint, explains why one JA4 digest can correspond to many HTTP User-Agent strings: JA4 hashes only TLS-layer fields (cipher suites, extension type IDs, SNI presence, ALPN) and ignores HTTP headers entirely.

    Score 0.58Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “JA4 Fingerprinting: Why One Digest Can Cover User-Agents”

    Comments on “JA4 Fingerprinting: Why One Digest Can Cover User-Agents”
  95. No. 95Score 0.56

    Optimizing your CI for agents in .NET Projects

    Summary and score for “Optimizing your CI for agents in .NET Projects”

    A post on .NET CI for monorepos explains why path filters fail when projects share code, and introduces dotnet-affected, a tool that uses the MSBuild project graph and a git diff to find which projects a change affects.

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Optimizing your CI for agents in .NET Projects”

    Comments on “Optimizing your CI for agents in .NET Projects”
  96. No. 96Score 0.58

    TrustGraph – ACLs that follow the data through your AI agents

    Summary and score for “TrustGraph – ACLs that follow the data through your AI agents”

    TrustGraph is an experimental open-source tool that enforces policy-as-code on what AI agents produce, based on the data those outputs were derived from. It addresses a gap exposed by Invariant Labs' May 2025 demonstration, where an agent read private repositories and published their contents to a public pull request, with…

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “TrustGraph – ACLs that follow the data through your AI agents”

    Comments on “TrustGraph – ACLs that follow the data through your AI agents”
  97. No. 97Score 0.60

    Some Linux maintainers are quitting under AI bugs, but there are signs of hope

    Summary and score for “Some Linux maintainers are quitting under AI bugs, but there are signs of hope”

    ZDNET reports that AI-driven bug scanning is flooding Linux and Windows maintainers with reports, with the Linux CVE team averaging 27 CVEs a day in 2026 and Windows Patch Tuesday listing 726 CVEs.

    Score 0.60Contestedjudges differ by 0.58 on overclaim

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Some Linux maintainers are quitting under AI bugs, but there are signs of hope”

    Comments on “Some Linux maintainers are quitting under AI bugs, but there are signs of hope”
  98. No. 98Score 0.58

    BrowserPod 4.0: the MicroVM that runs in the browser, now free for any use

    Summary and score for “BrowserPod 4.0: the MicroVM that runs in the browser, now free for any use”

    BrowserPod 4.0, from Leaning Technologies, makes its Linux-like WebAssembly microVM free to embed in any website via npm, with no account, API key, metering or telemetry.

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “BrowserPod 4.0: the MicroVM that runs in the browser, now free for any use”

    Comments on “BrowserPod 4.0: the MicroVM that runs in the browser, now free for any use”
  99. No. 99Score 0.58

    Ubuntu gets AI-generated security fixes for those feeling brave

    Summary and score for “Ubuntu gets AI-generated security fixes for those feeling brave”

    Canonical is offering Ubuntu Pro users an opt-in "Edge" archive of security patches generated by its Surge framework, a multi-agent system of LLMs that triages vulnerabilities, backports fixes, and tests them before escalating uncertain cases to staff.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “Ubuntu gets AI-generated security fixes for those feeling brave”

    Comments on “Ubuntu gets AI-generated security fixes for those feeling brave”
  100. No. 100Score 0.58

    Proof Automation in a Language

    Summary and score for “Proof Automation in a Language”

    This post surveys ingredients for future language design in the era of cheap automated proof: mutable value semantics, which avoids first-class references and their complexity (as in Swift and Hylo), restricted references that let tools like Aeneas skip separation logic, and design-by-contract with SMT-backed static checking (Dafny, F*, SPARK, Verus).…

    Score 0.58Contestedjudges differ by 0.42 on significance, 0.43 on practical relevance

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Proof Automation in a Language”

    Comments on “Proof Automation in a Language”
  101. No. 101Score 0.58

    U.S. data centers' combined electricity use rivals that of some states

    Summary and score for “U.S. data centers' combined electricity use rivals that of some states”

    A new study led by Harvard T.H. Chan School of Public Health, published in Nature Sustainability, estimates facility-level electricity use and carbon emissions for 403 U.S. hyperscale data centers from May 2024 to April 2025. It finds they used 68 to 99 terawatt-hours, roughly 1.8% of U.S.

    Score 0.58

    Elsewhere: HN 8 pts, not on front page

    Score breakdown for “U.S. data centers' combined electricity use rivals that of some states” · Also covered by WIRED, Harvard T.H. Chan School of Public Health

    Comments on “U.S. data centers' combined electricity use rivals that of some states”
  102. No. 102Score 0.57

    Interpretability-guided weight edits for Gemma 270M, no training

    Summary and score for “Interpretability-guided weight edits for Gemma 270M, no training”

    A Hacker News post by an individual developer describes making Google's 270-million-parameter Gemma 3 model better at following instructions without any gradient-based training, instead editing a few weights directly.

    Score 0.57Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Interpretability-guided weight edits for Gemma 270M, no training”

    Comments on “Interpretability-guided weight edits for Gemma 270M, no training”
  103. No. 103Score 0.57

    Footguns in Git CRLF conversions of text files

    Summary and score for “Footguns in Git CRLF conversions of text files”

    A developer, Bjørn Erling Fløtten, documents practical pitfalls in how Git converts CRLF and LF line endings in text files, tested on Git 2.53.0. The article's core advice is to run git add --renormalize after changing core.autocrlf or .gitattributes, and to heed Git's line-ending warnings, since index and working-tree states…

    Score 0.57

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Footguns in Git CRLF conversions of text files”

    Comments on “Footguns in Git CRLF conversions of text files”
  104. No. 104Score 0.57

    What's New in RubyGems/Bundler 4.1

    Summary and score for “What's New in RubyGems/Bundler 4.1”

    RubyGems and Bundler 4.1.0, published by the RubyGems team on the RubyGems blog, adds Gemfile-level override for rewriting dependency version constraints and prune for deleting download caches and git metadata after install.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “What's New in RubyGems/Bundler 4.1”

    Comments on “What's New in RubyGems/Bundler 4.1”
  105. No. 105Score 0.55

    AI chips could push data centers to the limit: 1,000 kW per rack

    Summary and score for “AI chips could push data centers to the limit: 1,000 kW per rack”

    Heise reports that AI workloads are pushing data center server racks from typical densities of 15 to 25 kW toward peaks of up to 1,000 kW by 2030, according to IDTechEx analyses.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “AI chips could push data centers to the limit: 1,000 kW per rack”

    Comments on “AI chips could push data centers to the limit: 1,000 kW per rack”
  106. No. 106Score 0.57

    Sagebrush is what I always wanted to write when I started SageMath

    Summary and score for “Sagebrush is what I always wanted to write when I started SageMath”

    William Stein, creator of SageMath, announces Sagebrush, an early-alpha, MIT/Apache-licensed rewrite of the most-used parts of Python, LaTeX, Magma, MATLAB and related tools, built in Rust and compiled to WebAssembly so it runs in a browser, on a desktop, or on a server.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Sagebrush is what I always wanted to write when I started SageMath”

    Comments on “Sagebrush is what I always wanted to write when I started SageMath”
  107. No. 107Score 0.57

    The small African lab sounding the alarm on AI bioweapons

    Summary and score for “The small African lab sounding the alarm on AI bioweapons”

    A Ghanaian AI lab, MinoHealth AI Labs, reported in a May 2025 arXiv preprint that after removing safeguards from its drug discovery model, the system generated designs for 1,020 proteins and 5,000 small molecules, some resembling the poison ricin, which predictive toxicity software scored as likely toxic.

    Score 0.57

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “The small African lab sounding the alarm on AI bioweapons”

    Comments on “The small African lab sounding the alarm on AI bioweapons”
  108. No. 108Score 0.57

    Tiny Voca: TTS on CPU (45M param, 18MB)

    Summary and score for “Tiny Voca: TTS on CPU (45M param, 18MB)”

    Muhammed Nazeem has released tiny-voca, an open-source English text-to-speech model with two voices and 4.5 million parameters, trained from scratch on about 18 hours of synthetic speech and running roughly 70 times faster than real time on four CPU threads.

    Score 0.57Contestedjudges differ by 0.43 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Tiny Voca: TTS on CPU (45M param, 18MB)”

    Comments on “Tiny Voca: TTS on CPU (45M param, 18MB)”
  109. No. 109Score 0.60

    Anthropic shortening its vulnerability disclosure timelines

    Summary and score for “Anthropic shortening its vulnerability disclosure timelines”

    Anthropic is cutting its coordinated vulnerability disclosure window from 90 days to 30 days for high- and critical-severity issues and to 45 days for others, applying to reports made from January 1, 2027.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Anthropic shortening its vulnerability disclosure timelines” · Also covered by The Next Web

    Comments on “Anthropic shortening its vulnerability disclosure timelines”
  110. No. 110Score 0.56

    S3rg: Grep for object storage, written in Rust

    Summary and score for “S3rg: Grep for object storage, written in Rust”

    S3rg is a command-line tool written in Rust that works like grep for object storage, searching S3, GCS and Azure prefixes or local paths with ripgrep's matching engine. Objects are streamed and searched without writing to disk, and gzip and zstd objects are decompressed on the fly.

    Score 0.56

    Elsewhere: HN 7 pts, not on front page

    Score breakdown for “S3rg: Grep for object storage, written in Rust”

    Comments on “S3rg: Grep for object storage, written in Rust”
  111. No. 111Score 0.56

    How Many Companies Are Using a Software Factory?

    Summary and score for “How Many Companies Are Using a Software Factory?”

    Public evidence suggests roughly 4,000 companies run "software factories," where coordinated agents handle issues through coding, review and merge with little human typing, though the author's central figure rests on undisclosed-share assumptions and could plausibly range from about 100 to 45,000.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “How Many Companies Are Using a Software Factory?”

    Comments on “How Many Companies Are Using a Software Factory?”
  112. No. 112Score 0.56

    Browser Immune System (BIS) Innate and Adaptive Threat Detection in the Browser

    Summary and score for “Browser Immune System (BIS) Innate and Adaptive Threat Detection in the Browser”

    The paper presents the Browser Immune System, a Chrome and Firefox extension that screens web pages in the browser using alias-web, a 4-billion-parameter cybersecurity transformer quantised to 4-bit weights and run on the user's GPU via WebGPU.

    Score 0.56Contestedjudges differ by 0.56 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Browser Immune System (BIS) Innate and Adaptive Threat Detection in the Browser”

    Comments on “Browser Immune System (BIS) Innate and Adaptive Threat Detection in the Browser”
  113. No. 113Score 0.56

    GenAI Is Erasing the Mid-Level Engineer Promo Signal

    Summary and score for “GenAI Is Erasing the Mid-Level Engineer Promo Signal”

    An engineering manager describes how generative AI is weakening a familiar promotion signal: design documents that once showed a mid-level engineer's reasoning now may not reveal whether the engineer independently navigated trade-offs or accepted a plausible model output.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “GenAI Is Erasing the Mid-Level Engineer Promo Signal”

    Comments on “GenAI Is Erasing the Mid-Level Engineer Promo Signal”
  114. No. 114Score 0.56

    Updates to Device Control and Data Access, and Full Disk Access in macOS

    Summary and score for “Updates to Device Control and Data Access, and Full Disk Access in macOS”

    Apple's developer post announces macOS 27.3 beta 2 changes to permission controls on Mac. Device Control and Data Access are split into System & App Protection Override and Accessibility, with per-app selection, more visibility into granted access, and a transparency indicator when one app controls another.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Updates to Device Control and Data Access, and Full Disk Access in macOS”

    Comments on “Updates to Device Control and Data Access, and Full Disk Access in macOS”
  115. No. 115Score 0.56

    How Billions in Safe Insurance Money Is Being Bet on A.I. Debt

    Summary and score for “How Billions in Safe Insurance Money Is Being Bet on A.I. Debt”

    Life insurers, which hold retirees' savings and are typically among the most conservative investors, now own $85.2 billion of AI debt, up 24 percent in six months and projected to reach up to $185 billion by next year.

    Score 0.56Contestedjudges differ by 0.41 on significance

    Elsewhere: HN 84 pts, front page

    Score breakdown for “How Billions in Safe Insurance Money Is Being Bet on A.I. Debt”

    Comments on “How Billions in Safe Insurance Money Is Being Bet on A.I. Debt”
  116. No. 116Score 0.56

    Benchmarking PDF-to-Markdown Tables (Docling, PyMuPDF, MarkItDown, and More)

    Summary and score for “Benchmarking PDF-to-Markdown Tables (Docling, PyMuPDF, MarkItDown, and More)”

    A benchmark published by Aura UG, the maker of the switchfile converter, compares PDF-to-Markdown tools on 82 public spreadsheets and Word documents rendered from LibreOffice, Word and Excel.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Benchmarking PDF-to-Markdown Tables (Docling, PyMuPDF, MarkItDown, and More)”

    Comments on “Benchmarking PDF-to-Markdown Tables (Docling, PyMuPDF, MarkItDown, and More)”
  117. No. 117Score 0.56

    W3C and the State of the Web in the AI Era

    Summary and score for “W3C and the State of the Web in the AI Era”

    The W3C Strategy Team's paper "W3C and the State of the Web in the AI Era," prepared by Philippe Le Hegaret and finalized on 20 October 2026, argues that AI agents have moved from reading the web to acting on it, so the web's standards agenda must adapt.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “W3C and the State of the Web in the AI Era”

    Comments on “W3C and the State of the Web in the AI Era”
  118. No. 118Score 0.56

    Japan suspends its Fugaku supercomputer after unauthorized access

    Summary and score for “Japan suspends its Fugaku supercomputer after unauthorized access”

    Japan's RIKEN institute has suspended its Fugaku supercomputer, built with Fujitsu at a cost of about 130 billion yen, after unauthorized access to the system used to log in and operate it, according to Kyodo News and the science ministry.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Japan suspends its Fugaku supercomputer after unauthorized access”

    Comments on “Japan suspends its Fugaku supercomputer after unauthorized access”
  119. No. 119Score 0.55

    Prefer synthetic traffic over LLM as a judge

    Summary and score for “Prefer synthetic traffic over LLM as a judge”

    The author argues that using LLMs as judges for search relevance labels merely bolts AI onto old evaluation metrics like NDCG, which ignore how the user interface shapes results.

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Prefer synthetic traffic over LLM as a judge”

    Comments on “Prefer synthetic traffic over LLM as a judge”
  120. No. 120Score 0.55

    Where Did the Flow Go?

    Summary and score for “Where Did the Flow Go?”

    Developers' sense that their work has grown more taxing and less enjoyable is attributed here to AI coding agents, which fragment attention with constant waiting and reviewing and remove the productive struggle that produces flow.

    Score 0.55Contestedjudges differ by 0.44 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Where Did the Flow Go?”

    Comments on “Where Did the Flow Go?”
  121. No. 121Score 0.55

    Claimseed, synthetic Medicare docs with planted errors and answer keys

    Summary and score for “Claimseed, synthetic Medicare docs with planted errors and answer keys”

    Claimseed is an open-source tool that generates fully synthetic U.S. medical-billing documents, including Medicare Summary Notices, itemized hospital bills, and Medicare Advantage EOBs, rendered as PDFs with ground-truth JSON listing the injected billing errors.

    Score 0.55Contestedjudges differ by 0.47 on reader interest, 0.42 on technical depth

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Claimseed, synthetic Medicare docs with planted errors and answer keys”

    Comments on “Claimseed, synthetic Medicare docs with planted errors and answer keys”
  122. No. 122Score 0.56

    Does weight-space programming need to touch the hidden layers?

    Summary and score for “Does weight-space programming need to touch the hidden layers?”

    An output-head experiment on Qwen3.5-0.8B edited only three output rows, leaving every transformer tensor unchanged, and scored 55.2% on MATH-500 against 52.2% for a published Astra-edited reference run that changed internal feed-forward layers.

    Score 0.56Contestedjudges differ by 0.42 on substantiated

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Does weight-space programming need to touch the hidden layers?”

    Comments on “Does weight-space programming need to touch the hidden layers?”

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 window1,381
Scored blind by two judges1,221
Off topic219
Read by the third judge238
Claim-checked25
Published122