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 89 stories · 6 hidden by your filters

  1. No. 1Score 0.81

    xAI ignoring Grokbot rootkit vulnerability

    Summary and score for “xAI ignoring Grokbot rootkit vulnerability”

    Security firm Profero reports that Grok Bot's cloud Linux sandbox can be taken over from the internet after one hostile command in an agent. Its chain starts with an unauthenticated Docker API that gives root on the VM. A hard-coded exec password ("local") plus a per-account token then allows remote shells. Profero says xAI and Cursor have not replied since August 24, 2026, and the findings held across 240+ builds. These are the researchers' own claims, unconfirmed independently. Builders should avoid shipping secrets in images.

    Score 0.81Claim check: 8 no source given

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “xAI ignoring Grokbot rootkit vulnerability”

    Comments on “xAI ignoring Grokbot rootkit vulnerability”
  2. No. 2Score 0.79

    Brotli CVE-2026-111497 reaches stock WordPress through pingbacks

    Summary and score for “Brotli CVE-2026-111497 reaches stock WordPress through pingbacks”

    Patchstack reports that CVE-2026-111497, an out-of-bounds write in Google's Brotli decoder, is reachable on a default WordPress install. An unauthenticated pingback to xmlrpc.php makes WordPress fetch an attacker's URL, and libcurl decodes the Brotli reply. Public tooling only crashes PHP workers, and no public chain gives code execution on PHP. The fix is the OS Brotli package, not a WordPress update; the wordpress:latest Docker image was still vulnerable on October 30. Several details, including the CVSS 9.2 score and CISA listing, were not independently verified.

    Score 0.79Claim check: 2 matches source · 5 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Brotli CVE-2026-111497 reaches stock WordPress through pingbacks”

    Comments on “Brotli CVE-2026-111497 reaches stock WordPress through pingbacks”
  3. No. 3Score 0.76

    Memory Safety Bugs in Rust Standard Library

    Summary and score for “Memory Safety Bugs in Rust Standard Library”

    Microsoft's FORGE Lab reported memory-safety bugs in Rust's core and std libraries, some triggerable from entirely safe code, found partly with its AI scanner Hyenas. They fall into three patterns: assuming code won't panic (MapWindows::clone), trusting user-written trait implementations (str::replace with a custom Pattern), and trusting process-wide OS state (process::exit, CommandExt::exec). The team rated practical exploitation of the signal-stack bug unlikely, an assessment limited to that issue. For builders, unsafe code must stay sound when caller-supplied code misbehaves. Some dates and fix details were not independently verified.

    Score 0.76Claim check: 4 matches source · 1 no source given · 3 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Memory Safety Bugs in Rust Standard Library”

    Comments on “Memory Safety Bugs in Rust Standard Library”
  4. No. 4Score 0.74

    31,088 secrets in Claude Code's .claude/settings.local.json on public GitHub

    Summary and score for “31,088 secrets in Claude Code's .claude/settings.local.json on public GitHub”

    GitGuardian reports that from January 2025 to November 7, 2026 it found 31,088 distinct secrets in Claude Code's .claude/settings.local.json on public GitHub, across 23,276 repositories, and 5,343 passed an automatic validity check at detection. The file fills up when users approve a command with "Yes, and don't ask again," which stores the command, key included. Secrets also reached 613 Docker Hub image secrets, 99 valid. These figures are GitGuardian's own and unverified independently. Anthropic's Boris Cherny said the team is investigating.

    Score 0.74Contestedjudges differ by 0.78 on new informationClaim check: 6 no source given · 2 couldn’t check

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “31,088 secrets in Claude Code's .claude/settings.local.json on public GitHub” · Also covered by GitGuardian Blog - Take Control of Your Secrets Security

    Comments on “31,088 secrets in Claude Code's .claude/settings.local.json on public GitHub”
  5. No. 5Score 0.70

    C# 15: Union Types, Closed Hierarchies, and the Unique One Value

    Summary and score for “C# 15: Union Types, Closed Hierarchies, and the Unique One Value”

    Christian Ekrem tried C# 15's new union keyword and closed classes on the .NET 11 release candidate. Switch expressions over a union are exhaustive, but a missing case only raises warning CS8509 unless made an error, and switch statements get no check. Because a union compiles to a struct, default holds a null that no case matches, so the documentation advises a null arm. He also found value types boxed, no built-in Result type, and System.Text.Json failing to deserialize unions. Several findings rest on his own runs.

    Score 0.70Claim 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 “C# 15: Union Types, Closed Hierarchies, and the Unique One Value”

    Comments on “C# 15: Union Types, Closed Hierarchies, and the Unique One Value”
  6. No. 6Score 0.69

    I Think, Therefore I Attack

    Summary and score for “I Think, Therefore I Attack”

    Researchers at UT Austin's Applied Research Laboratories tested "thought injection," where untrusted text mimics a reasoning model's own chain of thought. In a containerized cyber range, payloads written in the target's reasoning style and wrapped in thinking tags reached 50% success on repository backdooring and 42% on reverse-shell exfiltration against MiniMax-M2.5. The same instructions phrased as ordinary prompt injection never succeeded. Effects of tags varied by model, with 50 runs per cell on two models. Tag stripping looks unreliable, and filters should scan reasoning output too.

    Score 0.69Contestedjudges differ by 0.47 on substantiatedClaim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “I Think, Therefore I Attack”

    Comments on “I Think, Therefore I Attack”
  7. No. 7Score 0.68

    Claude Fable 5.2 and Mythos 5.2

    Summary and score for “Claude Fable 5.2 and Mythos 5.2”

    Anthropic released Claude Fable 5.2 and Mythos 5.2, the same underlying model, with Mythos limited to vetted cyber and life-science organizations. Pricing is $8 per million input tokens and $40 per million output tokens, with a fast mode at twice that price. On Anthropic's own benchmarks, Fable 5.2 scores 69.9% on FrontierSWE v2 against 62.3% for Opus 5.5. Anthropic's claims on speed, safeguard blocks, the Fermat formalization and sandbox-escape rates could not be independently checked, and the benchmark results come from the company's own evaluations.

    Score 0.68Contestedjudges differ by 0.74 on promotionalClaim check: 2 no source given · 6 couldn’t check

    Elsewhere: HN 969 pts, front page

    Score breakdown and claims for “Claude Fable 5.2 and Mythos 5.2” · Also covered by CNBC, ibm.com, anthropic.com, The Verge, CNBC, The Next Web

    Comments on “Claude Fable 5.2 and Mythos 5.2”
  8. No. 8Score 0.68

    What is malloc doing, anyway?

    Summary and score for “What is malloc doing, anyway?”

    Ben Andrew builds a toy malloc in stages (bump allocator, free list with splitting and coalescing, heap trimming, mmap for blocks of 128 KiB or more) and compares each with glibc. On a 500,000-step benchmark on one Xeon, glibc takes 0.0789 µs per step, while the free-list and trim versions run 65 and 142 times slower. The toy versions release memory better, but glibc's size bins and tcache buy speed at the cost of retained RSS. The figures are the author's own, from a single workload and machine.

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

    Elsewhere: HN 243 pts, front page

    Score breakdown and claims for “What is malloc doing, anyway?”

    Comments on “What is malloc doing, anyway?”
  9. No. 9Score 0.68

    XDG Base Directory Specification

    Summary and score for “XDG Base Directory Specification”

    The XDG Base Directory Specification, version 0.8 dated 8 May 2021, defines where applications should store user files through environment variables. Defaults are $HOME/.local/share for data, $HOME/.config for configuration, $HOME/.local/state for state and $HOME/.cache for cache. $XDG_RUNTIME_DIR must be user-owned with mode 0700, tied to login lifetime, and files need an access-time update every 6 hours or the sticky bit to avoid cleanup. The claim check found no independent sources, so these points rest on the specification text alone. For developers, it is the reference for placing app files predictably.

    Score 0.68Contestedjudges differ by 0.85 on new informationClaim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “XDG Base Directory Specification”

    Comments on “XDG Base Directory Specification”
  10. No. 10Score 0.67

    Phoenix vs. ClickHouse: Why the Fight Never Ends

    Summary and score for “Phoenix vs. ClickHouse: Why the Fight Never Ends”

    Alexander Novikov compares HBase (the storage under Phoenix) and ClickHouse on the same ten million documents, tracing their differences to four storage choices: index chunk size, read unit, update rewrite and key ownership. On his single-machine stand, a whole-row point read took 0.15 ms on HBase versus 3.47 ms on ClickHouse, while a three-column scan took 0.03 s versus 24.7 s. Single-row updates took 472 µs against 81.7 ms. The figures are his own measurements and are not independently verified.

    Score 0.67Contestedjudges differ by 0.42 on significanceClaim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Phoenix vs. ClickHouse: Why the Fight Never Ends”

    Comments on “Phoenix vs. ClickHouse: Why the Fight Never Ends”
  11. No. 11Score 0.71

    Home Assistant stores bearer tokens in cleartext; a password change revokes none

    Summary and score for “Home Assistant stores bearer tokens in cleartext; a password change revokes none”

    Researcher knell's "Not the Key" audit, summarized by Garrett Stimpson, reports that Home Assistant core 2026.11.0 keeps refresh-token secrets and JWT signing keys in cleartext in config/.storage/auth, and that none of the four password-change paths revokes tokens. System tokens and long-lived parents never expire. A live test on a disposable instance showed copied tokens still working after a password change. The claims rest on one audit and one local test, and no independent confirmation is cited. Builders of similar local credential stores should note the hashing and revocation gaps.

    Score 0.71Claim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Home Assistant stores bearer tokens in cleartext; a password change revokes none”

    Comments on “Home Assistant stores bearer tokens in cleartext; a password change revokes none”
  12. No. 12Score 0.67

    Score Centering: Understanding and Fixing Instability in RL

    Summary and score for “Score Centering: Understanding and Fixing Instability in RL”

    Cameron Wolfe's overview covers score centering, a method for stabilizing LLM reinforcement learning when the sampler and trainer disagree on token probabilities. Rather than importance sampling's multiplicative weights, it subtracts the sampler's expected score from the policy gradient, which the paper argues removes mismatch-induced drift. Qwen3-1.7B Countdown diagnostics show offline training is stable only with non-negative rewards, while online training is most stable with group-centered rewards. The headline gains, such as 30% versus 12% accuracy under severe quantization, have no independent source.

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

    Elsewhere: HN —

    Score breakdown and claims for “Score Centering: Understanding and Fixing Instability in RL”

    Comments on “Score Centering: Understanding and Fixing Instability in RL”
  13. No. 13Score 0.67

    MuJoCo 4: Studio and Filament, Deformables, Speculative Contacts

    Summary and score for “MuJoCo 4: Studio and Filament, Deformables, Speculative Contacts”

    Google DeepMind released MuJoCo 4.0, its open-source physics engine, with a new Filament-based Studio viewer that also runs in the browser, implicit deformables with plasticity and rods, and speculative contacts for thin or fast objects. The team reports that a 0.7 mm card is caught at a 16 ms timestep and that stair climbing and grasping run 6 to 9 times faster; these are the authors' own tests, and the speedup figures were not independently checked. Speculative contacts currently cover CPU only, and 3.x models need migration for soft bodies.

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

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “MuJoCo 4: Studio and Filament, Deformables, Speculative Contacts”

    Comments on “MuJoCo 4: Studio and Filament, Deformables, Speculative Contacts”
  14. No. 14Score 0.71

    unlocking any car from X brand in most of the world

    Summary and score for “unlocking any car from X brand in most of the world”

    A security researcher, writing with Sam Curry, describes how they found an exposed admin test-harness subdomain and a hardcoded OpenIDM admin credential in a major automaker's Android app and JavaScript bundles, which they used to reach customer account data and tokens across the automaker's connected car platform.

    Score 0.71Contestedjudges differ by 0.66 on overclaimClaim check: no checkable claims

    Elsewhere: HN 604 pts, front page

    Score breakdown and claims for “unlocking any car from X brand in most of the world”

    Comments on “unlocking any car from X brand in most of the world”
  15. No. 15Score 0.66

    Ernie Evolve: automated research system for improving AI training and inference

    Summary and score for “Ernie Evolve: automated research system for improving AI training and inference”

    Baidu describes ERNIE Evolve, an automated research system that proposes directions, edits code, and keeps only changes that pass a hidden evaluator and a code reviewer. It reports beating public bests on NanoGPT Speedrun Track 1 (38.9 s versus 39.9 s) and NanoChat AutoResearch, and scoring 0.819 on SOL-ExecBench against a leading 0.827. On ERNIE-5.5-Flash it claims a 33.7% GPU throughput gain with determinism preserved. All figures are Baidu's own; the Kunlunxin gain largely came from enabling existing optimized paths, and the leaderboard submission is unconfirmed.

    Score 0.66Claim 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 “Ernie Evolve: automated research system for improving AI training and inference” · Also covered by The Next Web

    Comments on “Ernie Evolve: automated research system for improving AI training and inference”
  16. No. 16Score 0.66

    Chrome 153 has a native MP4 tab recorder

    Summary and score for “Chrome 153 has a native MP4 tab recorder”

    Paul Irish's chrome-capture package uses Chrome's new Page.startScreenRecording DevTools method, which encodes AV1 into an mp4 inside the browser, with optional tab audio and no ffmpeg. That replaces the JPEG-over-CDP route behind Playwright and Puppeteer video. On one M4 Max replay, stop time fell from 3.8s to 17ms and WebSocket traffic from 10.5 MB to 450 KB. The figures are the author's own and unverified, from one workload; 120fps needs a 120Hz display and headful Chrome, and Windows is untested.

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

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Chrome 153 has a native MP4 tab recorder”

    Comments on “Chrome 153 has a native MP4 tab recorder”
  17. No. 17Score 0.63

    Remount sysfs to make its superblock rw: mount -O remount,rw /sys

    Summary and score for “Remount sysfs to make its superblock rw: mount -O remount,rw /sys”

    Engineer Peng Zhang traced a fourfold variance in Amazon EFS sequential read speed to a sysfs race. Containers in one network namespace share a single sysfs superblock, and whichever mounts first sets it ro or rw. If an unprivileged container goes first, efs-utils' privileged container can't write read_ahead_kb, leaving 128 KiB instead of 15 MiB. The proposed fix is mount -o remount,rw /sys. The kernel-behavior claims come from the author's source reading and a test script, and several weren't independently verified.

    Score 0.63Claim check: 2 matches source · 2 no source given · 4 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Remount sysfs to make its superblock rw: mount -O remount,rw /sys”

    Comments on “Remount sysfs to make its superblock rw: mount -O remount,rw /sys”
  18. No. 18Score 0.71

    On the hunt for ARTEX: understanding agentic attacks in AWS environments

    Summary and score for “On the hunt for ARTEX: understanding agentic attacks in AWS environments”

    Datadog Security Labs reports in-the-wild use of ARTEX, an AI-native penetration framework in which LLM agents plan and run attacks, against AWS environments in October. Observed paths were compromised long-term IAM keys, used to search CloudWatch logs for financial data and secrets, and public Cognito user pools, where the agent self-registered and tried to enumerate resources. Those discovery calls failed because the authenticated role lacked permissions. Datadog also found hundreds of exposed consoles. Defenders should rotate long-lived keys, restrict Cognito roles, and watch for AccessDenied bursts.

    Score 0.71Claim check: 3 matches source · 5 no source given

    Elsewhere: HN —

    Score breakdown and claims for “On the hunt for ARTEX: understanding agentic attacks in AWS environments”

    Comments on “On the hunt for ARTEX: understanding agentic attacks in AWS environments”
  19. No. 19Score 0.66

    Inference Performance Index: changing the way we measure inference capacity

    Summary and score for “Inference Performance Index: changing the way we measure inference capacity”

    Epoch AI introduced the Inference Performance Index, which scores chips by how many simultaneous agentic requests they can serve across 50–400 tokens/s/req, in GB300 equivalents. Its model puts GB300 at about 25× H100 capacity with 2.5× the FP8 FLOP/s, and H200 HGX at 0.12 and VR200 at 2.18. The scores are speed-of-light projections from four models and ignore software maturity. A validation on Kimi K3 matched GB300 measurements at 50% of modeled performance, within about 12%.

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

    Elsewhere: HN —

    Score breakdown and claims for “Inference Performance Index: changing the way we measure inference capacity”

    Comments on “Inference Performance Index: changing the way we measure inference capacity”
  20. No. 20Score 0.62

    Vgpu3d: Hardware 3D for Windows VM on QEMU/KVM, on the GPU your host has

    Summary and score for “Vgpu3d: Hardware 3D for Windows VM on QEMU/KVM, on the GPU your host has”

    vgpu3d is a Windows 11 WDDM driver that forwards a guest's 3D work through QEMU's virtio-gpu device to the host's Vulkan driver, giving KVM virtual machines hardware-accelerated Direct3D, OpenGL, Vulkan and OpenCL while the GPU remains shared with the host.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Vgpu3d: Hardware 3D for Windows VM on QEMU/KVM, on the GPU your host has”

    Comments on “Vgpu3d: Hardware 3D for Windows VM on QEMU/KVM, on the GPU your host has”
  21. No. 21Score 0.62

    Sandboxing on Amazon EKS with isolated and confidential pods

    Summary and score for “Sandboxing on Amazon EKS with isolated and confidential pods”

    AWS has added Nitro Sandboxes for EKS pods, ECS tasks and plain EC2. Each pod opts in through runtimeClassName and runs in its own Firecracker microVM with a Cedar admission policy. Isolated mode protects the node from the workload. Confidential mode treats the parent instance as untrusted and uses KMS-gated attestation. The documentation says vCPUs are time-shared and can be oversubscribed, so this is a security boundary, not a performance reservation. AWS says it costs nothing extra. The attestation and policy claims rest on AWS's own documentation.

    Score 0.62Claim check: 3 matches source · 3 no source given · 2 couldn’t check

    Elsewhere: HN 6 pts, not on front page

    Score breakdown and claims for “Sandboxing on Amazon EKS with isolated and confidential pods” · Also covered by Amazon Web Services, Inc., Amazon Web Services

    Comments on “Sandboxing on Amazon EKS with isolated and confidential pods”
  22. No. 22Score 0.70

    A Microphone Permission Bypass on iOS 26 (CVE-2026-65339)

    Summary and score for “A Microphone Permission Bypass on iOS 26 (CVE-2026-65339)”

    A researcher details CVE-2026-65339, a flaw in iOS 26 and related Apple systems that let a sandboxed app with no special entitlements record microphone audio after the user denies permission.

    Score 0.70Claim check: 2 no source given · 6 couldn’t check

    Elsewhere: HN 6 pts, not on front page

    Score breakdown and claims for “A Microphone Permission Bypass on iOS 26 (CVE-2026-65339)”

    Comments on “A Microphone Permission Bypass on iOS 26 (CVE-2026-65339)”
  23. No. 23Score 0.65

    What's new in Node.js 27

    Summary and score for “What's new in Node.js 27”

    The Node.js team published the first Alpha of Node.js 27 (v27.0.0-alpha.0.0.0) on October 29, 2026, under a new annual schedule that replaces the old October Current release with a six-month Alpha phase before a 27.0.0 Current release in April 2027.

    Score 0.65Claim check: 5 matches source · 1 not in source · 2 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “What's new in Node.js 27”

    Comments on “What's new in Node.js 27”
  24. No. 24Score 0.61

    EU to blacklist ten manipulative online design 'traps', leaked draft shows

    Summary and score for “EU to blacklist ten manipulative online design 'traps', leaked draft shows”

    According to a leaked draft of the Digital Fairness Act reported by Euronews, the European Commission plans to outright ban ten manipulative design practices, such as fake scarcity claims and pre-ticked add-ons, and bar paid loot boxes for minors. Fines could reach 6% of global turnover.

    Score 0.61Contestedjudges differ by 0.53 on reader interestClaim check: 1 matches source · 7 no source given

    Elsewhere: HN 17 pts, not on front page

    Score breakdown and claims for “EU to blacklist ten manipulative online design 'traps', leaked draft shows”

    Comments on “EU to blacklist ten manipulative online design 'traps', leaked draft shows”
  25. No. 25Score 0.60

    How SkyPilot on Slurm and Kubernetes put 1k idle GPU hours a week to work

    Summary and score for “How SkyPilot on Slurm and Kubernetes put 1k idle GPU hours a week to work”

    Photoroom's engineering team describes using SkyPilot to unify its on-premise GPUs, Slurm training cluster, and multi-provider Kubernetes inference fleet behind one task-file interface. Researchers can launch jobs from a single prompt, and agents route work to the right pool, including idle inference GPUs off peak.

    Score 0.60Claim check: 2 matches source · 5 no source given · 1 couldn’t check

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “How SkyPilot on Slurm and Kubernetes put 1k idle GPU hours a week to work”

    Comments on “How SkyPilot on Slurm and Kubernetes put 1k idle GPU hours a week to work”
  26. No. 26Score 0.69

    Attackers backdoor trusted AWS integrations to steal secrets

    Summary and score for “Attackers backdoor trusted AWS integrations to steal secrets”

    Attackers used compromised AWS IAM credentials tied to Qubole's account to assume cross-account integration roles in customer environments, then stole secrets and data from S3 buckets. Wiz, working with Mandiant, called it a broad, opportunistic supply-chain campaign affecting multiple customers.

    Score 0.69Claim check: 7 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Attackers backdoor trusted AWS integrations to steal secrets”

    Comments on “Attackers backdoor trusted AWS integrations to steal secrets”
  27. No. 27Score 0.65

    GR00T 2: Scaling robot foundation models for generalization, dexterity

    Summary and score for “GR00T 2: Scaling robot foundation models for generalization, dexterity”

    NVIDIA's GR00T 2 is a new generation of robot foundation models that jointly predict future video and robot actions, learning how physical interactions unfold. Trained on human egocentric video, handheld-gripper demonstrations, and robot teleoperation data, it transfers skills to robots that never practiced them, succeeding zero-shot on 40% of 50…

    Score 0.65

    Elsewhere: HN 394 pts, front page

    Score breakdown for “GR00T 2: Scaling robot foundation models for generalization, dexterity”

    Comments on “GR00T 2: Scaling robot foundation models for generalization, dexterity”
  28. No. 28Score 0.72

    AI coding agents leak credentials into public repositories

    Summary and score for “AI coding agents leak credentials into public repositories”

    Security firm Apiiro reports that a documented Claude Code option, "Yes, and don't ask again," can save live credentials into .claude/settings.local.json, and that the file sometimes escapes Git's ignore protection.

    Score 0.72Contestedjudges differ by 0.55 on new information

    Elsewhere: HN —

    Score breakdown for “AI coding agents leak credentials into public repositories” · Also covered by Ars Technica: AI

    Comments on “AI coding agents leak credentials into public repositories”
  29. No. 29Score 0.64

    Zebkit – Chromium's engine as a library you link, not a runtime you live in

    Summary and score for “Zebkit – Chromium's engine as a library you link, not a runtime you live in”

    Zebkit is a library that packages Chromium's 157 engine (Blink, V8, PDFium, FFmpeg and more) for direct linking into native applications, in C, C++, Rust, Python, Go or JavaScript, without Chrome's browser or Electron's runtime.

    Score 0.64Contestedjudges differ by 0.45 on substantiated

    Elsewhere: HN 112 pts, front page

    Score breakdown for “Zebkit – Chromium's engine as a library you link, not a runtime you live in”

    Comments on “Zebkit – Chromium's engine as a library you link, not a runtime you live in”
  30. No. 30Score 0.69

    Alpine Linux and OpenSSH patch flaw that's as bad as it can get

    Summary and score for “Alpine Linux and OpenSSH patch flaw that's as bad as it can get”

    A newly disclosed OpenSSH authentication bypass, CVE-2026-123158, lets attackers log into passwordless accounts on Alpine Linux and other musl-based distributions by sending a password between 257 and 1,024 bytes long. The flaw stems from musl's crypt() returning "*" on error, which matches the password field of key-only accounts.

    Score 0.69Contestedjudges differ by 0.55 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Alpine Linux and OpenSSH patch flaw that's as bad as it can get”

    Comments on “Alpine Linux and OpenSSH patch flaw that's as bad as it can get”
  31. No. 31Score 0.65

    Open-source app that lets LLMs control your iPhone (no Mac required)

    Summary and score for “Open-source app that lets LLMs control your iPhone (no Mac required)”

    An open-source iOS app, iphone-use-app, lets an LLM control the iPhone it runs on, with no Mac needed at run time and no extra hardware or jailbreak. A Mac is only required once to create a pairing file.

    Score 0.65

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Open-source app that lets LLMs control your iPhone (no Mac required)”

    Comments on “Open-source app that lets LLMs control your iPhone (no Mac required)”
  32. No. 32Score 0.57

    UX Roundup: Human “Taste” Less Unique | World Usability Day | New Work With AI | Agent Waits | Pale Text | AI Teammates | Evidence | Proximity Compatibility Principle | Back Button

    Summary and score for “UX Roundup: Human “Taste” Less Unique | World Usability Day | New Work With AI | Agent Waits | Pale Text | AI Teammates | Evidence | Proximity Compatibility Principle | Back Button”

    A market experiment with 1,025 consumers, reported by researchers John Conlon and Peter Schwardmann, found that AI-designed mini games beat amateur games by 27% and outsold professionals' games, and that disclosing human authorship did not raise willingness to pay.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “UX Roundup: Human “Taste” Less Unique | World Usability Day | New Work With AI | Agent Waits | Pale Text | AI Teammates | Evidence | Proximity Compatibility Principle | Back Button”

    Comments on “UX Roundup: Human “Taste” Less Unique | World Usability Day | New Work With AI | Agent Waits | Pale Text | AI Teammates | Evidence | Proximity Compatibility Principle | Back Button”
  33. No. 33Score 0.65

    Raw RAG leaks 0%, but agent memory leaks 67%

    Summary and score for “Raw RAG leaks 0%, but agent memory leaks 67%”

    A Show HN project released a deterministic benchmark and linter for access-control leaks in agent memory. Its tests on a 24-document corpus found that vector stores enforce document permissions correctly for raw chunk retrieval (0% leaks), but background synthesis jobs drop parent ACLs, leaking restricted content in 2 of 3…

    Score 0.65Contestedjudges differ by 0.44 on substantiated

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Raw RAG leaks 0%, but agent memory leaks 67%”

    Comments on “Raw RAG leaks 0%, but agent memory leaks 67%”
  34. No. 34Score 0.65

    Airgap-check – find what your AI app silently downloads offline

    Summary and score for “Airgap-check – find what your AI app silently downloads offline”

    airgap-check is an open-source tool that finds what AI applications try to download or contact when run without network access, which matters for air-gapped, hospital, defense, or no-internet deployments.

    Score 0.65

    Elsewhere: HN 7 pts, not on front page

    Score breakdown for “Airgap-check – find what your AI app silently downloads offline”

    Comments on “Airgap-check – find what your AI app silently downloads offline”
  35. No. 35Score 0.57

    The Race to Build a "Find My" for AI Chips

    Summary and score for “The Race to Build a "Find My" for AI Chips”

    A bill passed by the U.S. House Foreign Affairs Committee would require location verification for advanced AI chips, so that export restrictions can be enforced after chips leave the factory.

    Score 0.57

    Elsewhere: HN 23 pts, front page

    Score breakdown for “The Race to Build a "Find My" for AI Chips”

    Comments on “The Race to Build a "Find My" for AI Chips”
  36. No. 36Score 0.65

    Do AI coding agents care about SemVer or is it something for humans only?

    Summary and score for “Do AI coding agents care about SemVer or is it something for humans only?”

    A Microsoft Developer Advocate's write-up, published on the Microsoft developer blog, reports experiments showing that AI coding agents sometimes treat semantic version labels as signals about risk.

    Score 0.65

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Do AI coding agents care about SemVer or is it something for humans only?”

    Comments on “Do AI coding agents care about SemVer or is it something for humans only?”
  37. No. 37Score 0.65

    Researchers pwn automakers’ Android apps to remotely unlock cars, honk horns

    Summary and score for “Researchers pwn automakers’ Android apps to remotely unlock cars, honk horns”

    Security researchers Shubham Shah of Assetnote and Sam Curry found flaws in the Android apps of two major automakers, using weaknesses in the ForgeRock identity platform to gain remote control of connected cars in Australia and the US. They could start engines, flash headlights, and sound horns.

    Score 0.65Contestedjudges differ by 0.85 on new information

    Elsewhere: HN —

    Score breakdown for “Researchers pwn automakers’ Android apps to remotely unlock cars, honk horns”

    Comments on “Researchers pwn automakers’ Android apps to remotely unlock cars, honk horns”
  38. No. 38Score 0.65

    A 9B with an agentic search loop beat a 35B MoE with one-shot retrieval

    Summary and score for “A 9B with an agentic search loop beat a 35B MoE with one-shot retrieval”

    A deep-dive measures agentic search against one-shot retrieval on 14 unseen official reports, using the author's own app, Matrix. A dense 9B with agentic search found 65% of facts, beating a 35B MoE with one-shot retrieval (48%).

    Score 0.65

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A 9B with an agentic search loop beat a 35B MoE with one-shot retrieval”

    Comments on “A 9B with an agentic search loop beat a 35B MoE with one-shot retrieval”
  39. No. 39Score 0.64

    Your Experts Have Partners. There's HOPE for Keeping Them Together

    Summary and score for “Your Experts Have Partners. There's HOPE for Keeping Them Together”

    HOPE is a second-order method for pruning experts from Mixture-of-Experts language models, described in a paper by its authors. Existing pruning methods score each expert independently, but HOPE accounts for pairwise interactions between co-activated experts, which matters most at aggressive pruning rates.

    Score 0.64

    Elsewhere: HN —

    Score breakdown for “Your Experts Have Partners. There's HOPE for Keeping Them Together”

    Comments on “Your Experts Have Partners. There's HOPE for Keeping Them Together”
  40. No. 40Score 0.56

    Coordinated disturbances haunt grids worldwide

    Summary and score for “Coordinated disturbances haunt grids worldwide”

    Researchers led by Dirk Witthaut at Forschungszentrum Jülich traced recurring once-a-minute frequency sags on grids from Europe to Australia to uninterruptible power supplies and grid-scale batteries that check their charge state on the same minute.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Coordinated disturbances haunt grids worldwide”

    Comments on “Coordinated disturbances haunt grids worldwide”
  41. No. 41Score 0.56

    I reverse-engineered my Pixel Watch's raw PPG sensor to check for sleep apnea

    Summary and score for “I reverse-engineered my Pixel Watch's raw PPG sensor to check for sleep apnea”

    A developer tested whether a Pixel Watch 1's raw optical pulse signal and motion data could help check for sleep apnea after a Fitbit showed blood oxygen dipping to 80–87% for an hour.

    Score 0.56Contestedjudges differ by 0.48 on substantiated

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “I reverse-engineered my Pixel Watch's raw PPG sensor to check for sleep apnea”

    Comments on “I reverse-engineered my Pixel Watch's raw PPG sensor to check for sleep apnea”
  42. No. 42Score 0.64

    E-means: 10x faster k-means at equal loss

    Summary and score for “E-means: 10x faster k-means at equal loss”

    E-means is a mini-batch k-means method that replaces Lloyd's full-batch updates with exponential moving averages of centroid sums and counts, plus cosine-scheduled step sizes, dead-cluster revival and a finalization stage.

    Score 0.64

    Elsewhere: HN 331 pts, front page

    Score breakdown for “E-means: 10x faster k-means at equal loss”

    Comments on “E-means: 10x faster k-means at equal loss”
  43. No. 43Score 0.56

    Number of /64 subnets and ASes affected by routing loops and amplification

    Summary and score for “Number of /64 subnets and ASes affected by routing loops and amplification”

    A measurement dataset published on an IPv6 subnet-router anycast page reports 413 million /64 subnets triggering routing loops, 779 thousand triggering amplification, and 6.5 thousand ASes affected by loops. The likely cause is customers using default routes on provider-assigned address space without NULL routes for unused subnets.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Number of /64 subnets and ASes affected by routing loops and amplification”

    Comments on “Number of /64 subnets and ASes affected by routing loops and amplification”
  44. No. 44Score 0.56

    Making a €40 USB-C hub work as a Nintendo Switch 2 dock

    Summary and score for “Making a €40 USB-C hub work as a Nintendo Switch 2 dock”

    A hobbyist documents how they made a €40 USB-C hub act as a Nintendo Switch 2 dock by reverse-engineering the console's proprietary USB-PD vendor messages with a sniffer, patching the hub's VL103 PD controller firmware to answer the handshake, and then tracking a washed-out HDR picture to the RTD2142 DisplayPort-to-HDMI…

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Making a €40 USB-C hub work as a Nintendo Switch 2 dock”

    Comments on “Making a €40 USB-C hub work as a Nintendo Switch 2 dock”
  45. No. 45Score 0.63

    Fable 5.2 and Jagged AGI

    Summary and score for “Fable 5.2 and Jagged AGI”

    Ethan Mollick's post on Substack argues that Anthropic's newest model, Claude Fable 5.2, is "almost" artificial general intelligence, which he calls "Jagged AGI," because it excels at demanding tasks like building games, running a simulated business plan, reconstructing a literature course, and drafting academic research while still showing uneven judgment…

    Score 0.63

    Elsewhere: HN 8 pts, not on front page

    Score breakdown for “Fable 5.2 and Jagged AGI”

    Comments on “Fable 5.2 and Jagged AGI”
  46. No. 46Score 0.63

    The Agents on Rails Files: enter lemans, a Ruby harness for fair agent races

    Summary and score for “The Agents on Rails Files: enter lemans, a Ruby harness for fair agent races”

    lemans is an open source Ruby harness from the Rails Foundation and Evil Martians for running fair, sandboxed races between coding agents. Tasks are a patch that plants a bug, a reference solution, and a hidden Ruby test that grades the result.

    Score 0.63Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The Agents on Rails Files: enter lemans, a Ruby harness for fair agent races”

    Comments on “The Agents on Rails Files: enter lemans, a Ruby harness for fair agent races”
  47. No. 47Score 0.63

    ChatGPT Desktop forwards your SSH agent to remote machines

    Summary and score for “ChatGPT Desktop forwards your SSH agent to remote machines”

    A security researcher reports that ChatGPT Desktop forwards the user's local SSH agent to remote machines, letting an agent running there authenticate with the user's SSH credentials. That could let it push to private repositories such as those on GitHub.

    Score 0.63

    Elsewhere: HN 207 pts, front page

    Score breakdown for “ChatGPT Desktop forwards your SSH agent to remote machines”

    Comments on “ChatGPT Desktop forwards your SSH agent to remote machines”
  48. No. 48Score 0.63

    Google Gemini routes personal Drive files and emails to human annotators

    Summary and score for “Google Gemini routes personal Drive files and emails to human annotators”

    Google has reportedly reversed its 2023 promise that personal Gmail, Docs, and Drive content connected to Gemini is not seen by human reviewers, with updated help documentation saying connected app data may be routed into human review.

    Score 0.63Contestedjudges differ by 0.41 on overclaim

    Elsewhere: HN 506 pts, front page

    Score breakdown for “Google Gemini routes personal Drive files and emails to human annotators”

    Comments on “Google Gemini routes personal Drive files and emails to human annotators”
  49. No. 49Score 0.63

    Extending LLDB to Debug Applications That Span CPUs and GPUs

    Summary and score for “Extending LLDB to Debug Applications That Span CPUs and GPUs”

    Meta describes a vendor-neutral accelerator plugin architecture for LLDB's lldb-server that lets one debugging session cover both CPU and GPU code. It represents the CPU and GPU as separate debugger targets presented together, with hardware-specific backends translating vendor APIs into common debugger concepts.

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “Extending LLDB to Debug Applications That Span CPUs and GPUs”

    Comments on “Extending LLDB to Debug Applications That Span CPUs and GPUs”
  50. No. 50Score 0.63

    G-harness-evals: Which agent harness does tasks best, for the least cost?

    Summary and score for “G-harness-evals: Which agent harness does tasks best, for the least cost?”

    A benchmark by Garry Tan's g-harness-evals project compares agent harnesses (Claude Code, Codex, Hermes Agent, OpenClaw, and QM's Pi loop) on 20 public multi-user tasks, holding the model fixed at Opus 5.5 or GPT-6.1 Sol.

    Score 0.63

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “G-harness-evals: Which agent harness does tasks best, for the least cost?”

    Comments on “G-harness-evals: Which agent harness does tasks best, for the least cost?”
  51. No. 51Score 0.63

    Fireworks AI October 2026 security incident

    Summary and score for “Fireworks AI October 2026 security incident”

    Fireworks AI, an AI inference platform, has updated its October 2026 security bulletin after an intrusion that the company links to a financially motivated group practicing "LLMJacking," which resells stolen LLM access.

    Score 0.63Contestedjudges differ by 0.42 on new information

    Elsewhere: HN 7 pts, not on front page

    Score breakdown for “Fireworks AI October 2026 security incident” · Also covered by Bloomberg.com

    Comments on “Fireworks AI October 2026 security incident”
  52. No. 52Score 0.62

    AI labs raise prices without raising prices

    Summary and score for “AI labs raise prices without raising prices”

    Playcode's analysis reports that AI labs are keeping per-token prices flat while raising effective costs by increasing output tokens. OpenAI's GPT Image 2.6 used 3.3x the output tokens of 2.5 on identical prompts, and Anthropic's tokenizer change makes the same text about 30% longer.

    Score 0.62

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI labs raise prices without raising prices”

    Comments on “AI labs raise prices without raising prices”
  53. No. 53Score 0.62

    Run Linux applications in hypervisor-isolated micro-VMs with Hyperlight

    Summary and score for “Run Linux applications in hypervisor-isolated micro-VMs with Hyperlight”

    Hyperlight, a CNCF Sandbox project, is a library virtual machine monitor that embeds micro-VMs isolated by the host's hypervisor into applications, aiming to close the gap between fast but weakly isolated sandboxes and slow full VMs.

    Score 0.62

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Run Linux applications in hypervisor-isolated micro-VMs with Hyperlight”

    Comments on “Run Linux applications in hypervisor-isolated micro-VMs with Hyperlight”
  54. No. 54Score 0.62

    The Average Is Free. Difference Is Not

    Summary and score for “The Average Is Free. Difference Is Not”

    A tech essay argues that AI models converge on one "winning" answer, so everyone's first drafts look alike, and it cites studies showing the sameness: a NeurIPS paper where 1,250 attempts produced a few repeated metaphors, a German study where ChatGPT's jokes were mostly 25 recycled ones, and Wharton research…

    Score 0.62Contestedjudges differ by 0.46 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “The Average Is Free. Difference Is Not”

    Comments on “The Average Is Free. Difference Is Not”
  55. No. 55Score 0.64

    Hackers exploit SSHole OpenSSH flaw to breach Linux systems

    Summary and score for “Hackers exploit SSHole OpenSSH flaw to breach Linux systems”

    Attackers are actively exploiting "SSHole" (CVE-2026-123158), a flaw in OpenSSH on musl-based Linux systems such as Alpine, disclosed by researcher Ozgun Kultekin. An oversized password makes musl's hashing function return "*", which OpenSSH wrongly accepts for accounts with disabled passwords, enabling root access.

    Score 0.64Contestedjudges differ by 0.50 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Hackers exploit SSHole OpenSSH flaw to breach Linux systems” · Also covered by Ars Technica: IT

    Comments on “Hackers exploit SSHole OpenSSH flaw to breach Linux systems”
  56. No. 56Score 0.61

    Pain of Coding vs. Pain of Configuring

    Summary and score for “Pain of Coding vs. Pain of Configuring”

    Coding has become cheap enough that enterprises should reconsider the long-standing build-versus-buy default, which favored purchasing vendor software and configuring it. The author argues agents shift the crossover point toward building, since a company only needs to serve its own narrow requirements and can operate software through agent-driven tooling.

    Score 0.61Contestedjudges differ by 0.41 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Pain of Coding vs. Pain of Configuring”

    Comments on “Pain of Coding vs. Pain of Configuring”
  57. No. 57Score 0.60

    Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here

    Summary and score for “Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here”

    Samir Sarwar has published a self-taught video series and C++17/CUDA project, samir-vllm, that builds a small LLM inference engine for Llama 3.2 1B from scratch. It covers safetensors weight loading, a BPE tokenizer, transformer layers, a paged KV cache, and static or continuous batching, with companion handwritten notes.

    Score 0.60Contestedjudges differ by 0.47 on substantiated

    Elsewhere: HN 77 pts, front page

    Score breakdown for “Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here”

    Comments on “Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here”
  58. No. 58Score 0.60

    A Critique of GPUI

    Summary and score for “A Critique of GPUI”

    In "A Critique of GPUI," the author argues that GPUI and egui's state models are dead ends for complex desktop apps, because any UI framework must know what changed and who depends on it to scale. GPUI's notify call carries only an entity id, so it re-renders most views.

    Score 0.60Contestedjudges differ by 0.48 on substantiated

    Elsewhere: HN 137 pts, front page

    Score breakdown for “A Critique of GPUI”

    Comments on “A Critique of GPUI”
  59. No. 59Score 0.60

    JAX single controller used to train Gemini is now GA on GCP

    Summary and score for “JAX single controller used to train Gemini is now GA on GCP”

    Google's developer blog announces general availability of Pathways on Google Cloud Platform, the distributed system Google uses to train Gemini. Pathways uses a single Python client that describes the whole computation as a graph, while a distributed runtime places it, moves tensors, and keeps accelerators busy.

    Score 0.60

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “JAX single controller used to train Gemini is now GA on GCP”

    Comments on “JAX single controller used to train Gemini is now GA on GCP”
  60. No. 60Score 0.55

    Quick thoughts on the GCP europe-west2-b outage

    Summary and score for “Quick thoughts on the GCP europe-west2-b outage”

    A Google Cloud zonal outage in europe-west2-b on October 19 was caused by a cable fault on one power feed, combined with an unrelated maintenance breaker left open on the other feed and a PLC configuration that never triggered the backup diesel generators.

    Score 0.55Contestedjudges differ by 0.55 on new information

    Elsewhere: HN —

    Score breakdown for “Quick thoughts on the GCP europe-west2-b outage”

    Comments on “Quick thoughts on the GCP europe-west2-b outage”
  61. No. 61Score 0.60

    Keep the Agent's Packets Out of the Host Kernel

    Summary and score for “Keep the Agent's Packets Out of the Host Kernel”

    A blog post by Kunal Kushwaha, published 9 November 2026, examines a Firecracker VM escape reported on 7 November, in which a guest exploits virtio checksum-offload hints to make the host kernel read and write its memory, then hijacks a root-run VMM.

    Score 0.60Contestedjudges differ by 0.41 on promotional

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Keep the Agent's Packets Out of the Host Kernel”

    Comments on “Keep the Agent's Packets Out of the Host Kernel”
  62. No. 62Score 0.60

    500 Lines or Less: A Simple Object Model

    Summary and score for “500 Lines or Less: A Simple Object Model”

    Carl Friedrich Bolz's chapter "500 Lines or Less: A Simple Object Model" builds a simplified Python-style object model in Python, starting from classes, instances, and attribute reads and writes.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “500 Lines or Less: A Simple Object Model”

    Comments on “500 Lines or Less: A Simple Object Model”
  63. No. 63Score 0.60

    Beyond Mvn Install: Maven Practices That Pay Off

    Summary and score for “Beyond Mvn Install: Maven Practices That Pay Off”

    The article offers practical Maven 3.9 build techniques drawn from open-source Java projects maintained by its author, including a defaultGoal set to verify, a fast-build profile that skips slow steps, and property-based version management. It also introduces the basepom parent POMs, which supply pinned, sensible plugin defaults.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Beyond Mvn Install: Maven Practices That Pay Off”

    Comments on “Beyond Mvn Install: Maven Practices That Pay Off”
  64. No. 64Score 0.60

    Cargo Takes Minutes. Mitts Takes Seconds: 12 Rustls Worktrees

    Summary and score for “Cargo Takes Minutes. Mitts Takes Seconds: 12 Rustls Worktrees”

    A study comparing Cargo, Cargo with sccache, and Mitts on parallel builds of 100 open-source Rust projects found Mitts finishing a twelve-worktree rustls build in 9.3 seconds versus Cargo's 3 minutes 42 seconds, with prewarmed shared caches.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Cargo Takes Minutes. Mitts Takes Seconds: 12 Rustls Worktrees”

    Comments on “Cargo Takes Minutes. Mitts Takes Seconds: 12 Rustls Worktrees”
  65. No. 65Score 0.60

    Black Ops II statically recompiled to native ARM64 macOS

    Summary and score for “Black Ops II statically recompiled to native ARM64 macOS”

    A developer has built bo2-mac, an unofficial tool that statically recompiles the Windows executables of Black Ops II to native ARM64 for Apple M1 Macs, using Remill to lift x86 code and a hand-written Win32 runtime with Direct3D 11 over Metal.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Black Ops II statically recompiled to native ARM64 macOS”

    Comments on “Black Ops II statically recompiled to native ARM64 macOS”
  66. No. 66Score 0.59

    The race to secure record-like information for domains

    Summary and score for “The race to secure record-like information for domains”

    The post argues that securing simple key-value configuration for domain names is still an unsolved problem, since DNSSEC is uncommon (adoption ranges from about 1% to 63% depending on cohort) and HTTPS-based .well-known files don't fit every protocol.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “The race to secure record-like information for domains”

    Comments on “The race to secure record-like information for domains”
  67. No. 67Score 0.59

    The sound that can give away where you live

    Summary and score for “The sound that can give away where you live”

    An article by Ryan Porter and various bots, published by an unnamed outlet, argues that phone recordings can leak location through clues like the Doppler shift of passing aircraft or thunder matched to lightning records.

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The sound that can give away where you live”

    Comments on “The sound that can give away where you live”
  68. No. 68Score 0.59

    Private LLM copies can agree on a secret key over a public channel

    Summary and score for “Private LLM copies can agree on a secret key over a public channel”

    Researchers report that identical LLM copies agree on random binary preference questions more than they agree with other models, even across labs. This similarity lets two copies derive a shared secret key over a public channel using one-way communication and enables fingerprinting to confirm two instances are the same model.…

    Score 0.59Contestedjudges differ by 0.42 on substantiated

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Private LLM copies can agree on a secret key over a public channel”

    Comments on “Private LLM copies can agree on a secret key over a public channel”
  69. No. 69Score 0.59

    Let’s Talk About Why all This OT Stuff is Connected

    Summary and score for “Let’s Talk About Why all This OT Stuff is Connected”

    The author, an industrial incident responder, argues that exposed OT systems are not simply the result of incompetence. Their explanation covers IT/OT convergence, uniform vendor architectures that make attacks scalable, and economic pressure toward remote monitoring and support.

    Score 0.59

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Let’s Talk About Why all This OT Stuff is Connected”

    Comments on “Let’s Talk About Why all This OT Stuff is Connected”
  70. No. 70Score 0.58

    Hot fridges, a fake nuclear alert, a $6B heist: dawn of a new hacking epidemic?

    Summary and score for “Hot fridges, a fake nuclear alert, a $6B heist: dawn of a new hacking epidemic?”

    A Guardian report describes a spate of cyber-attacks in late October and early November 2026: hackers hijacked commercial refrigeration systems across several European countries, pushing temperatures as high as 50C, while a $6.3bn crypto heist hit Binance and a fake nuclear alert was sent to people in Kyiv.

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Hot fridges, a fake nuclear alert, a $6B heist: dawn of a new hacking epidemic?”

    Comments on “Hot fridges, a fake nuclear alert, a $6B heist: dawn of a new hacking epidemic?”
  71. No. 71Score 0.58

    Where does Baumol bite? · Casey Handmer

    Summary and score for “Where does Baumol bite? · Casey Handmer”

    Casey Handmer argues that the largest private capital build-out in history, roughly $720–745 billion of 2026 hyperscaler capex mostly for AI infrastructure, will be limited less by intelligence than by inputs that resist cost declines.

    Score 0.58Contestedjudges differ by 0.42 on technical depth, 0.62 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Where does Baumol bite? · Casey Handmer”

    Comments on “Where does Baumol bite? · Casey Handmer”
  72. No. 72Score 0.58

    I ran 120B MoE models on my 32GB M1 Max using under 16 GB

    Summary and score for “I ran 120B MoE models on my 32GB M1 Max using under 16 GB”

    A developer ran 120-billion-parameter mixture-of-experts models on a 32GB M1 Max by keeping most expert weights on SSD under a 16 GiB process-memory budget. Their engine averaged 7.84 tokens/s versus 5.78 for TUFF 10.2.2 on matched BF16 weights, and degraded less when another process used memory.

    Score 0.58Contestedjudges differ by 0.41 on substantiated

    Elsewhere: HN 24 pts, not on front page

    Score breakdown for “I ran 120B MoE models on my 32GB M1 Max using under 16 GB”

    Comments on “I ran 120B MoE models on my 32GB M1 Max using under 16 GB”
  73. No. 73Score 0.58

    Attack warning: SonicWall SSLVPN flaw is under attack

    Summary and score for “Attack warning: SonicWall SSLVPN flaw is under attack”

    Attackers are actively targeting a flaw in the SSLVPN feature of SonicOS on SonicWall firewalls, which lets users with valid credentials for another login method bypass SAML authentication and, apparently, the multifactor check it is meant to enforce (CVE-2026-19868).

    Score 0.58Contestedjudges differ by 0.41 on overclaim

    Elsewhere: HN —

    Score breakdown for “Attack warning: SonicWall SSLVPN flaw is under attack”

    Comments on “Attack warning: SonicWall SSLVPN flaw is under attack”
  74. No. 75Score 0.57

    Nvidia’s GR00T 2 teaches robots delicate skills from human video

    Summary and score for “Nvidia’s GR00T 2 teaches robots delicate skills from human video”

    Nvidia has released GR00T 2, an open "world action model" that helps robots learn delicate hand skills partly from human video, including 100,000 hours of first-person footage it is making public.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “Nvidia’s GR00T 2 teaches robots delicate skills from human video”

    Comments on “Nvidia’s GR00T 2 teaches robots delicate skills from human video”
  75. No. 76Score 0.57

    How we made table parsing 4× faster with DFlash

    Summary and score for “How we made table parsing 4× faster with DFlash”

    Extend.ai reports that fine-tuning a DFlash speculative-decoding drafter on its own table-parsing model cut median Light Table latency on one H100 from 3.35 to 0.76 seconds, about 4.4× faster. The gain comes from raising acceptance length from 6.73 to 12.48 tokens per verification pass on Performance Table.

    Score 0.57

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “How we made table parsing 4× faster with DFlash”

    Comments on “How we made table parsing 4× faster with DFlash”
  76. No. 77Score 0.57

    PolinRider: North Korea-linked worm forges developers' commits in 7,609 repos

    Summary and score for “PolinRider: North Korea-linked worm forges developers' commits in 7,609 repos”

    A North Korea-linked malware campaign, dubbed PolinRider, has reportedly infected 7,609 code repositories by forging developer commits. The worm's spread through software projects threatens supply-chain trust, since maliciously altered code can appear to come from legitimate contributors. Developers and companies should audit repository histories and commit signing practices.

    Score 0.57Contestedjudges differ by 0.71 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “PolinRider: North Korea-linked worm forges developers' commits in 7,609 repos”

    Comments on “PolinRider: North Korea-linked worm forges developers' commits in 7,609 repos”
  77. No. 78Score 0.57

    AI Is Causing Signal Collapse Everywhere All at Once

    Summary and score for “AI Is Causing Signal Collapse Everywhere All at Once”

    The author argues that AI makes it cheap to fake the signals people use as shortcuts for judging quality, including GitHub stars, polished writing, homework answers, and related-work sections, causing many signals to fail at once.

    Score 0.57

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “AI Is Causing Signal Collapse Everywhere All at Once”

    Comments on “AI Is Causing Signal Collapse Everywhere All at Once”
  78. No. 79Score 0.58

    Particle-1 vs. GPT-6.1 SOL: same pass rate, 71% lower cost in our tests

    Summary and score for “Particle-1 vs. GPT-6.1 SOL: same pass rate, 71% lower cost in our tests”

    In a 90-trial controlled test inside Groovy's execution system, Particle-1 High matched GPT-6.1 SOL's 28 passes out of 30 targeted and data-science trials while costing about 71% less at peak provider rates, including failed attempts. Particle-1 Max passed 26.

    Score 0.58Contestedjudges differ by 0.40 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Particle-1 vs. GPT-6.1 SOL: same pass rate, 71% lower cost in our tests”

    Comments on “Particle-1 vs. GPT-6.1 SOL: same pass rate, 71% lower cost in our tests”
  79. No. 80Score 0.57

    FairPlay Has Been Cracked

    Summary and score for “FairPlay Has Been Cracked”

    UnfairPlay is a standalone Go tool, published by its authors, that recovers a FairPlay model from an enrolled Mac and uses it to decrypt personalized iOS IPAs, including converting them for macOS.

    Score 0.57

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “FairPlay Has Been Cracked”

    Comments on “FairPlay Has Been Cracked”
  80. No. 81Score 0.56

    We brought ViT-H coordination below a 16 GB memory budget

    Summary and score for “We brought ViT-H coordination below a 16 GB memory budget”

    A systems engineering write-up by Denis Voronko reports that coordinating a full ViT-H/14 fine-tuning run peaked at 9.6 GB of process RAM, down from 29.4 GB, on a shared MI300X test setup.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “We brought ViT-H coordination below a 16 GB memory budget”

    Comments on “We brought ViT-H coordination below a 16 GB memory budget”
  81. No. 82Score 0.56

    The DOs and don’ts of post-quantum DNSSEC

    Summary and score for “The DOs and don’ts of post-quantum DNSSEC”

    A piece on post-quantum DNSSEC explains why cryptographically relevant quantum computers threaten DNSSEC key pairs, noting that a shorter key lifetime helps only so far. It compares ECDSA P-256 with the post-quantum ML-DSA-44 algorithm, showing that ML-DSA-44 signatures are far larger, which inflates DNS responses and matters for operators planning…

    Score 0.56Contestedjudges differ by 0.48 on new information

    Elsewhere: HN —

    Score breakdown for “The DOs and don’ts of post-quantum DNSSEC”

    Comments on “The DOs and don’ts of post-quantum DNSSEC”
  82. No. 83Score 0.56

    A Chromium Bug Caused by a 25-Hour Day

    Summary and score for “A Chromium Bug Caused by a 25-Hour Day”

    A Chromium bug, traced to a code path that assumed every day lasts 24 hours, caused a scheduled metrics task to fire repeatedly on the 25-hour daylight saving "fall back" day in regions like Vancouver.

    Score 0.56Contestedjudges differ by 0.46 on audience breadth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A Chromium Bug Caused by a 25-Hour Day”

    Comments on “A Chromium Bug Caused by a 25-Hour Day”
  83. No. 84Score 0.59

    AI R&D Is Accelerating Towards Potential RSI in 2027

    Summary and score for “AI R&D Is Accelerating Towards Potential RSI in 2027”

    METR published a report estimating that AI is already accelerating AI R&D at Anthropic by roughly 2x, with capability progress since October 2025 running about twice the pace of the prior period (wide confidence interval of roughly 1.3x to 3x).

    Score 0.59Contestedjudges differ by 0.57 on new information, 0.43 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI R&D Is Accelerating Towards Potential RSI in 2027”

    Comments on “AI R&D Is Accelerating Towards Potential RSI in 2027”
  84. No. 85Score 0.56

    Breaking a Rust actor cluster on purpose, and what it taught us

    Summary and score for “Breaking a Rust actor cluster on purpose, and what it taught us”

    The author describes a Rust actor cluster demo for the tellus project, where a verifier reported zero violations during a 45-minute chaos run yet 8 of 5,352 requests failed.

    Score 0.56Contestedjudges differ by 0.42 on reader interest

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Breaking a Rust actor cluster on purpose, and what it taught us”

    Comments on “Breaking a Rust actor cluster on purpose, and what it taught us”
  85. No. 86Score 0.56

    We shrank DSV4-Pro by 24%, within 0.5pt of original on MMLU – Tileward

    Summary and score for “We shrank DSV4-Pro by 24%, within 0.5pt of original on MMLU – Tileward”

    Tileward's Anantha Sharma reports a compressed build of DeepSeek-V4-Pro, called TWX3, that is 24% smaller (675.2 GB against 892.7 GB) while scoring 89.6% on MMLU against the published model's 90.0%, within the half-point margin set beforehand.

    Score 0.56Contestedjudges differ by 0.43 on substantiated, 0.50 on promotional, 0.48 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “We shrank DSV4-Pro by 24%, within 0.5pt of original on MMLU – Tileward”

    Comments on “We shrank DSV4-Pro by 24%, within 0.5pt of original on MMLU – Tileward”
  86. No. 87Score 0.55

    Security Vocabulary for Maintainers

    Summary and score for “Security Vocabulary for Maintainers”

    The article is a glossary for open source maintainers that separates the vulnerability-handling terms: a report is the private tip, an advisory is the public document you publish after a fix, a CVE ID is a unique identifier, a CVE Record is the public database entry, and a CNA is…

    Score 0.55Contestedjudges differ by 0.44 on technical depth, 0.66 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Security Vocabulary for Maintainers”

    Comments on “Security Vocabulary for Maintainers”
  87. No. 88Score 0.55

    You've Given the Insider Threats in Your Company a Super Intelligent Accomplice

    Summary and score for “You've Given the Insider Threats in Your Company a Super Intelligent Accomplice”

    An opinion essay titled "You've Given the Insider Threats in Your Company a Super Intelligent Accomplice" argues that enterprise AI assistants, which are being connected to email, Jira, Confluence, infrastructure records, and security data, consolidate organizational knowledge that could make a single malicious or financially pressured employee far more damaging.…

    Score 0.55Contestedjudges differ by 0.48 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “You've Given the Insider Threats in Your Company a Super Intelligent Accomplice”

    Comments on “You've Given the Insider Threats in Your Company a Super Intelligent Accomplice”
  88. No. 89Score 0.55

    Courses built from Hacker News posts

    Summary and score for “Courses built from Hacker News posts”

    Hackertutorials.com is a catalog of eleven self-study courses, such as building a database, a compiler, or an operating system, each assembled from tutorials that Hacker News users upvoted.

    Score 0.55Contestedjudges differ by 0.43 on audience breadth

    Elsewhere: HN 11 pts, not on front page

    Score breakdown for “Courses built from Hacker News posts”

    Comments on “Courses built from Hacker News posts”
  89. No. 90Score 0.55

    Linus Torvalds sees signs AI-prompted patch flood has peaked

    Summary and score for “Linus Torvalds sees signs AI-prompted patch flood has peaked”

    Linus Torvalds, the Linux kernel's lead developer, says in his weekly update on Linux 7.4 rc2 that AI-assisted bug fixes are arriving in volume, and that the eBPF subsystem's maintainer, Alexei Starovoitov, reports fixes now outpacing new reports there.

    Score 0.55Contestedjudges differ by 0.49 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Linus Torvalds sees signs AI-prompted patch flood has peaked”

    Comments on “Linus Torvalds sees signs AI-prompted patch flood has peaked”
  90. No. 91Score 0.55

    AI Bubble Tracker

    Summary and score for “AI Bubble Tracker”

    The AI Bubble Tracker is a dashboard of charts comparing AI spending, revenue, debt, adoption and prices against past boom-and-bust cycles, drawing on SEC filings, FRED, Epoch AI and other public sources.

    Score 0.55Contestedjudges differ by 0.43 on new information

    Elsewhere: HN 8 pts, front page

    Score breakdown for “AI Bubble Tracker”

    Comments on “AI Bubble Tracker”
  91. No. 92Score 0.55

    AI learns from the products of thinking, not thought itself

    Summary and score for “AI learns from the products of thinking, not thought itself”

    A paper by Kallen, Alhasan, Crone, and Kaplan argues that AI systems trained on finished outputs, such as paintings, recordings, and papers, learn only the traces of expertise and miss the trial-and-error process behind them.

    Score 0.55

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “AI learns from the products of thinking, not thought itself”

    Comments on “AI learns from the products of thinking, not thought itself”
  92. No. 93Score 0.55

    Vercel MCP Now Supports Client ID Metadata Documents

    Summary and score for “Vercel MCP Now Supports Client ID Metadata Documents”

    Vercel's MCP server now accepts connections from any MCP client that publishes a Client ID Metadata Document, a JSON file hosted on its own domain whose HTTPS URL serves as the OAuth client_id, so clients no longer need manual review.

    Score 0.55

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Vercel MCP Now Supports Client ID Metadata Documents”

    Comments on “Vercel MCP Now Supports Client ID Metadata Documents”
  93. No. 94Score 0.55

    Before an AI agent touches your money

    Summary and score for “Before an AI agent touches your money”

    Twilio's CIO argues that AI agents moving into finance workflows like invoice matching and payment routing need the same controls people have: separate verified identities, explicit approval scopes, tamper-proof audit logs, and clean handoffs to humans.

    Score 0.55Contestedjudges differ by 0.46 on new information

    Elsewhere: HN —

    Score breakdown for “Before an AI agent touches your money”

    Comments on “Before an AI agent touches your money”
  94. No. 95Score 0.58

    Supply chain attack on Qubole customers

    Summary and score for “Supply chain attack on Qubole customers”

    Sentra, working with Wiz and Mandiant, reported a supply chain attack in which an attacker stole credentials belonging to Qubole, a big-data platform, and used cross-account roles that dozens of its customers had granted to quietly pull data from their cloud environments.

    Score 0.58Contestedjudges differ by 0.70 on promotional, 0.45 on overclaim

    Elsewhere: HN 17 pts, front page

    Score breakdown for “Supply chain attack on Qubole customers”

    Comments on “Supply chain attack on Qubole customers”

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 window1,127
Scored blind by two judges1,027
Off topic164
Read by the third judge235
Claim-checked24
Published95