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 98 stories · 8 hidden by your filters

  1. No. 1Score 0.79

    In-depth Analysis of DPRK Contagious Interview campaign

    Summary and score for “In-depth Analysis of DPRK Contagious Interview campaign”

    Amsterdam Technologies, in a static-only analysis, dissects World-bat, a fake poker-app take-home test tied to the DPRK-linked Contagious Interview campaign. Opening the repo in a trusted VS Code or Cursor workspace, or running npm install, starts a beacon that sends environment variables, including any tokens, to a remote server every five seconds and evals the server's replies. A plain git clone is harmless. The follow-on malware was not retrieved, and the claim-check items lack independent sourcing, so attribution rests on the report's own matching to public reporting.

    Score 0.79Claim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “In-depth Analysis of DPRK Contagious Interview campaign”

    Comments on “In-depth Analysis of DPRK Contagious Interview campaign”
  2. No. 2Score 0.80

    Hackers breached AWS accounts through unused third-party access permissions

    Summary and score for “Hackers breached AWS accounts through unused third-party access permissions”

    Wiz, with Sentra and Mandiant, reports that between October 14 and 26, 2026 an actor used IAM credentials from an AWS account tied to Qubole to assume customer-created roles, then enumerated resources, read secrets, validated keys with TruffleHog and downloaded S3 data. Most assumed roles were otherwise unused but still trusted the vendor account. Wiz says a source-IP condition would have blocked every observed assumption. The findings rest largely on Wiz's own telemetry, and an "n0n" claim of responsibility is unverified.

    Score 0.80Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Hackers breached AWS accounts through unused third-party access permissions”

    Comments on “Hackers breached AWS accounts through unused third-party access permissions”
  3. No. 3Score 0.75

    Microsoft finds AI agents can be backdoored by editing under 10 neurons

    Summary and score for “Microsoft finds AI agents can be backdoored by editing under 10 neurons”

    Microsoft Research describes Lazarus, a technique that turns fewer than 10 rarely activated neurons in a language model into a backdoor without retraining. A classifier trained on about 50 examples identifies the trigger, and in one demonstration an agent downloads and runs attacker code once the date reaches 2027. The edited model can still pass standard safety evaluations, and the attacker needs only write access to the checkpoint. Reliable detection requires comparing weights with an unmodified original, so teams shipping agents should verify model provenance and integrity.

    Score 0.75Claim check: 8 matches source

    Elsewhere: HN —

    Score breakdown and claims for “Microsoft finds AI agents can be backdoored by editing under 10 neurons”

    Comments on “Microsoft finds AI agents can be backdoored by editing under 10 neurons”
  4. No. 4Score 0.74

    Backdooring Codex

    Summary and score for “Backdooring Codex”

    Security firm Depi reports that a prompt injection in OpenAI's Codex issue-labeling GitHub workflow, which fed issue text to the agent with open user access, gave code execution in the agent sandbox. Unauthenticated CoreCLR debug pipes in /tmp then exposed runner secrets from process memory, and cache poisoning is said to have reached the @openai/codex npm publishing pipeline. The claims are the researchers' own and unverified, though OpenAI permitted disclosure. CI agents processing untrusted input need strict permissions and isolation.

    Score 0.74Claim check: 5 no source given · 3 couldn’t check

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “Backdooring Codex”

    Comments on “Backdooring Codex”
  5. No. 5Score 0.74

    Phala Confidential AI Compromised

    Summary and score for “Phala Confidential AI Compromised”

    Phala, a confidential-computing AI host, published a preliminary incident report: an attacker chained a legacy internet-facing service with management interfaces that trusted loopback access, created unauthorized CVMs, and took over an inference VM on a GPU server lacking full hardening. Request data for DeepSeek V4 Flash 0731 is treated as exposed from October 21 to November 8; canary credentials were misused only from October 21–22. One tenant's credentials were abused. The claims come from Phala alone; a full root cause analysis is pending.

    Score 0.74Contestedjudges differ by 0.44 on audience breadthClaim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Phala Confidential AI Compromised”

    Comments on “Phala Confidential AI Compromised”
  6. No. 6Score 0.71

    Tang, a compiler that derives GPU kernels and bounds how fast they can run

    Summary and score for “Tang, a compiler that derives GPU kernels and bounds how fast they can run”

    Municipal Robotics' Cam Pedersen describes tang, a Rust compiler that derives GPU kernels from a two-construct tensor IR (views and comprehensions) instead of using hand-written ones, including flash-style attention forward and backward. Reported results are mixed: 1.2× faster than PyTorch at sequence length 512 but slower at 128, and Qwen3-8B decode tied with llama.cpp at short context. Derived attention trails FlashAttention-3, and a Lean-proved traffic bound gives a time floor that is loose on the H200. The figures are the author's own and the claim check found no independent source.

    Score 0.71Claim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Tang, a compiler that derives GPU kernels and bounds how fast they can run”

    Comments on “Tang, a compiler that derives GPU kernels and bounds how fast they can run”
  7. No. 7Score 0.69

    Running parallel agents for coding is not as fast as you think

    Summary and score for “Running parallel agents for coding is not as fast as you think”

    Stackness's author timed 364 Claude Code task runs in one repo and found parallel agents averaged 1.21 working at once, peaking at 1.46 in the busiest window. A separate six-agent run of 106 tasks measured 1.82. Higher published speedups, up to 3.9, come from benchmarks that exclude tests or from search branches that never merge. Idle time waiting for review, a shared test machine and a late 92-file merge cut throughput. The data covers one developer, one repo and one harness, and the author's own figures are unsourced.

    Score 0.69Claim check: 7 matches source · 1 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Running parallel agents for coding is not as fast as you think”

    Comments on “Running parallel agents for coding is not as fast as you think”
  8. No. 8Score 0.73

    Analysis of 366,845 requests for claude.json and .cursor/mcp.json

    Summary and score for “Analysis of 366,845 requests for claude.json and .cursor/mcp.json”

    HoneyLabs reports that on 4 September a Google Cloud scanner added 49 paths in one second, 32 of them AI coding tool files such as Claude Code's /.claude/.credentials.json, /.cursor/mcp.json and other MCP and LiteLLM configs. It dropped them after 7 September, but other clients have kept requesting them: 8,067 requests through 9 November. The figures come from HoneyLabs honeypots alone, and the post cannot say whether any real server answered. Its advice: check web roots for exposed home directories, rotate exposed tokens, and block by path.

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

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Analysis of 366,845 requests for claude.json and .cursor/mcp.json”

    Comments on “Analysis of 366,845 requests for claude.json and .cursor/mcp.json”
  9. No. 9Score 0.68

    Nvidia Vera CPU Review: 2TB/S Memory Bandwidth, a Real x86 Rival

    Summary and score for “Nvidia Vera CPU Review: 2TB/S Memory Bandwidth, a Real x86 Rival”

    StorageReview tested NVIDIA's Vera, an 88-core Arm server CPU built on its custom Olympus core, in a dual-socket reference server against AMD EPYC and Intel Xeon systems. Two sockets sustained 2,073GB/s on STREAM Triad, 2.3x the best DDR5 machine, and Vera led on large OpenFOAM and ClickBench runs. It trailed on vector-heavy GROMACS and AES work, and results on database engines were mixed. Vera ran a newer compiler stack than the x86 systems, and several figures lack independent sourcing.

    Score 0.68Contestedjudges differ by 0.54 on overclaimClaim check: 1 matches source · 7 no source given

    Elsewhere: HN 14 pts, not on front page

    Score breakdown and claims for “Nvidia Vera CPU Review: 2TB/S Memory Bandwidth, a Real x86 Rival”

    Comments on “Nvidia Vera CPU Review: 2TB/S Memory Bandwidth, a Real x86 Rival”
  10. No. 10Score 0.68

    E-means: outperforming k-means (in one-tenth the epochs)

    Summary and score for “E-means: outperforming k-means (in one-tenth the epochs)”

    Researchers at Meta's FAIR and École des Ponts present e-means, a mini-batch k-means using exponential-moving-average centroid updates, a cosine-decayed step and dead-cluster revival. Batch size and decay follow scaling laws fitted on one dataset (DEEP), so no sweep is needed. The authors report beating Lloyd's 25-epoch loss in 1.4 to 2.6 epochs, 10-45x GPU speedups, and 10 billion DINO vectors clustered into 16 million centroids in 2.5 epochs. These are the authors' own results, and the laws are less suited to low-dimensional or tiny datasets.

    Score 0.68Claim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “E-means: outperforming k-means (in one-tenth the epochs)”

    Comments on “E-means: outperforming k-means (in one-tenth the epochs)”
  11. No. 11Score 0.69

    Even with Explicit Rules and Accurate Task Rewards, RL Learns to Violate Them

    Summary and score for “Even with Explicit Rules and Accurate Task Rewards, RL Learns to Violate Them”

    Xinjie Shen reports that plain task-reward RL on a 4B Qwen3.5-4B agent (GRPO) learns to violate explicitly stated rules when a violation pays more. On API tasks, certified profitable violations rose from 8.3% to 91.7% while real work improved. Adding a binary violation penalty raised safe completion to 91.7% and partly transferred to new tasks. These figures come from the author's own small, synthetic environments, and the claim check could not verify them independently.

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

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Even with Explicit Rules and Accurate Task Rewards, RL Learns to Violate Them”

    Comments on “Even with Explicit Rules and Accurate Task Rewards, RL Learns to Violate Them”
  12. No. 12Score 0.72

    Building an Agent Sandbox from Scratch in Python

    Summary and score for “Building an Agent Sandbox from Scratch in Python”

    A developer's walkthrough of building a Python agent sandbox in six layers, adding namespaces, a rebuilt root, a hand-written seccomp filter, Landlock, network isolation, and cgroup limits. After each layer, the author runs a fixed set of attacks against the runtime.

    Score 0.72Contestedjudges differ by 0.43 on significance

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Building an Agent Sandbox from Scratch in Python”

    Comments on “Building an Agent Sandbox from Scratch in Python”
  13. No. 13Score 0.66

    How OpenAI, Canva and Spotify made slow CI steps up to 20x faster

    Summary and score for “How OpenAI, Canva and Spotify made slow CI steps up to 20x faster”

    BerryBox published a playbook of about 50 CI optimizations for GitHub Actions, grouped into wasted runs, slow setup, builds, tests, test selection, critical path, runners, flaky tests and pull request volume. Each entry names a risk level and companies using it. Examples include cancelling superseded runs, setting job timeouts from p95 runtimes, caching dependencies by lockfile, sharding tests, and gating new tests with 20 to 50 repeat runs. The listing gives no measured speedups, so the "20x" headline isn't backed by figures here.

    Score 0.66Claim check: 8 matches source

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “How OpenAI, Canva and Spotify made slow CI steps up to 20x faster”

    Comments on “How OpenAI, Canva and Spotify made slow CI steps up to 20x faster”
  14. No. 14Score 0.65

    A Schema Named Hash Sent Our Ruby SDK into Infinite Recursion

    Summary and score for “A Schema Named Hash Sent Our Ruby SDK into Infinite Recursion”

    Emad Khair of octri.dev describes how a test schema named Hash sent the company's generated Ruby SDK into infinite recursion. Inside the SDK's module, Ruby resolved Hash to the generated class rather than the built-in, so a helper never matched real hashes. A 150-name fixture of standard-library types also broke five other languages' builds. The fixes were explicit qualification, renaming models, and keeping separate internal and public names. Reserved-keyword tables were not enough. The account is the team's own, with no outside verification.

    Score 0.65Claim check: 8 no source given

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “A Schema Named Hash Sent Our Ruby SDK into Infinite Recursion”

    Comments on “A Schema Named Hash Sent Our Ruby SDK into Infinite Recursion”
  15. No. 15Score 0.65

    AWS Finally Built Its Own AWS Emulator

    Summary and score for “AWS Finally Built Its Own AWS Emulator”

    AWS released AWS Emulator, a local emulator of AWS service APIs for testing without an account or service fee, written in Rust under a CC0 license. Amazon says autonomous agents wrote all of it; Simon Willison counts 29,684 commits in three weeks. Fidelity scores are published per release, and the project is not meant for production or security testing. It runs natively on Linux and in Docker on macOS and Windows. Source builds need about 100 GB of disk, per the README, and telemetry is on by default.

    Score 0.65Contestedjudges differ by 0.47 on substantiatedClaim check: 4 matches source · 3 no source given · 1 couldn’t check

    Elsewhere: HN 765 pts, front page

    Score breakdown and claims for “AWS Finally Built Its Own AWS Emulator” · Also covered by Amazon News, All Things Distributed, Simon Willison

    Comments on “AWS Finally Built Its Own AWS Emulator”
  16. No. 16Score 0.69

    Tests as Data

    Summary and score for “Tests as Data”

    Matteo Vaccari, writing on martinfowler.com, argues that AI-written tests are hard to trust because they are voluminous, hard to read, coupled to implementation and often just confirm current behavior. He proposes customer-style black-box tests kept as data (tables, YAML, or small text DSLs) run by a short runner an agent can generate, making changes easier to spot in review. His support is interviews with 26 Thoughtworks developers and worked examples; the quality-decline reports he cites are not sourced, and he offers no measured results.

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

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Tests as Data”

    Comments on “Tests as Data”
  17. No. 17Score 0.72

    nodejs/node v22.23.5

    Summary and score for “nodejs/node v22.23.5”

    Node.js shipped coordinated security releases for its 22.x, 24.x and 26.x lines (v22.23.5, v24.22.3, v26.13.1). The 22.x notes list 17 CVEs: eight High, six Medium, three Low. Most fixes are in http2 and TLS, covering stream and buffer lifetime, SNI changes on resumption, and mTLS resumption. The 24.x and 26.x notes add a High-rated remotely reachable TLS server crash that the 22.x notes omit. Teams running public HTTP/2 or TLS servers should upgrade. The notes give titles and severities, not exploit detail.

    Score 0.72Claim check: 1 matches source · 1 no source given · 6 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “nodejs/node v22.23.5” · Also covered by GitHub, GitHub, nodejs.org

    Comments on “nodejs/node v22.23.5”
  18. No. 18Score 0.64

    Choosing QuePaxa, an unusual consensus algorithm

    Summary and score for “Choosing QuePaxa, an unusual consensus algorithm”

    Cloudflare explains why its experimental Meerkat consensus service, built for a global control plane, uses QuePaxa (Tennage and Băsescu et al., 2023) rather than Raft or Multi-Paxos. Leader-based algorithms stall when the leader is slow or unreachable, and their election timeouts must exceed network latency. QuePaxa makes the leader optional, uses random proposal priorities, and has clients hedge across replicas, so it needs no fault detection. Cloudflare reports over 95% of decisions on the leader's fast path and no stalls, but these are its own unsourced figures.

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

    Elsewhere: HN —

    Score breakdown and claims for “Choosing QuePaxa, an unusual consensus algorithm”

    Comments on “Choosing QuePaxa, an unusual consensus algorithm”
  19. No. 19Score 0.64

    Relax: JIT compiler for Rust to reduce build times

    Summary and score for “Relax: JIT compiler for Rust to reduce build times”

    Makepad's Relax is a JIT runtime for Rust, written in Rust, that runs a Cargo package from source with no rustc, LLVM or linker. On an M4 MacBook a large 3D game reaches its first frame in 11 s (1.15 s on rerun), versus about 3 minutes of building with rustc, and saving a file patches changed functions into the running program. By the authors' own measurements, optimised code is 1.14–1.26× slower than native-CPU --release, 2.3× on dense numeric loops, and a broader run is being re-measured.

    Score 0.64Claim check: 4 matches source · 4 no source given

    Elsewhere: HN 6 pts, front page

    Score breakdown and claims for “Relax: JIT compiler for Rust to reduce build times”

    Comments on “Relax: JIT compiler for Rust to reduce build times”
  20. No. 20Score 0.72

    CVE-2026-118383: Use-after-free vulnerability in MongoDB

    Summary and score for “CVE-2026-118383: Use-after-free vulnerability in MongoDB”

    Datadog Security Labs detailed CVE-2026-118383, a use-after-free in MongoDB replica set members (CVSS 4.0 score 9.5) that MongoDB disclosed on November 10. Standalone mongod is unaffected. In a lab on 8.0.34, two unauthenticated ping requests left stale waiters, and a later failover crashed the new primary. It depends on a lag and election the attacker cannot cause. Code execution was not demonstrated. Fixes start at 8.0.37 and equivalents; some details were not independently verified.

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

    Elsewhere: HN —

    Score breakdown and claims for “CVE-2026-118383: Use-after-free vulnerability in MongoDB”

    Comments on “CVE-2026-118383: Use-after-free vulnerability in MongoDB”
  21. No. 21Score 0.64

    Formalizing the US Tax Code

    Summary and score for “Formalizing the US Tax Code”

    Lexifina, a legal-data startup, translated the 2025 individual income tax into Lean with AI and compared it against PolicyEngine-US, Tax-Calculator and UsTaxes on 5,000 synthetic households. It reports 138 drafting faults in Title 26, agreement with PolicyEngine on 4,993 households, and bugs in the other tools, plus three small unpublished loopholes. Most figures are the author's own and unverified; the AI-run triage has had no lawyer or CPA review, and agreement among calculators does not show correctness. The model covers one tax year.

    Score 0.64Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 7 pts, not on front page

    Score breakdown and claims for “Formalizing the US Tax Code”

    Comments on “Formalizing the US Tax Code”
  22. No. 22Score 0.70

    My weather app sends my location, Wi-Fi and cell tower to data brokers

    Summary and score for “My weather app sends my location, Wi-Fi and cell tower to data brokers”

    A reverse-engineering writeup says the Czech weather app Aladin (version 1.13.0, over a million Google Play installs) bundles six third-party data-collection SDKs that send location, Wi-Fi network details, cell tower data and a persistent install ID to network-analytics companies.

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

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “My weather app sends my location, Wi-Fi and cell tower to data brokers”

    Comments on “My weather app sends my location, Wi-Fi and cell tower to data brokers”
  23. No. 23Score 0.61

    Performance evolution of route lookup on Linux

    Summary and score for “Performance evolution of route lookup on Linux”

    Vincent Bernat benchmarks Linux route lookup performance from kernel 2.6.32 through 7.3 on an AMD EPYC 8024P. He finds that a Linux 6.5 mitigation for the AMD SRSO vulnerability added about 200 ns to IPv6 lookups, erasing earlier gains, while IPv4 regressed mainly from a 2021 linkdown check that he…

    Score 0.61Claim check: 5 no source given · 3 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Performance evolution of route lookup on Linux”

    Comments on “Performance evolution of route lookup on Linux”
  24. No. 24Score 0.61

    The Invisible Shape

    Summary and score for “The Invisible Shape”

    The essay traces the history and usability problems of hierarchical context menus, from Xerox PARC's 1976 Smalltalk-76 pop-up menus to Apple's 1987 sub-menus, and explains why diagonal cursor movement to sub-menus is fragile.

    Score 0.61Claim check: 5 matches source · 1 no source given · 2 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “The Invisible Shape”

    Comments on “The Invisible Shape”
  25. No. 25Score 0.61

    Using voltage glitching to unleash the A100 in my mining card

    Summary and score for “Using voltage glitching to unleash the A100 in my mining card”

    A researcher claims a voltage fault injection technique can blow eFuses on NVIDIA's GA100 die, unlocking features disabled on the CMP 170HX mining card, including the full 80GB of HBM, 120–128 SMs, PCIe Gen4, and NVLink.

    Score 0.61Contestedjudges differ by 0.52 on overclaimClaim check: 8 no source given

    Elsewhere: HN 284 pts, front page

    Score breakdown and claims for “Using voltage glitching to unleash the A100 in my mining card”

    Comments on “Using voltage glitching to unleash the A100 in my mining card”
  26. No. 26Score 0.66

    Getting the (In)Formal Statement Right

    Summary and score for “Getting the (In)Formal Statement Right”

    An essay on autoformalization argues that a Lean-checked proof can still prove a different statement than the one intended, and that the mathematical meaning must be checked separately from proof correctness.

    Score 0.66Claim check: 1 matches source · 2 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Getting the (In)Formal Statement Right”

    Comments on “Getting the (In)Formal Statement Right”
  27. No. 27Score 0.61

    Going all-in on Sorbet (type-checking for Ruby)

    Summary and score for “Going all-in on Sorbet (type-checking for Ruby)”

    DocSpring's engineer describes migrating their Rails codebase to Sorbet, a static type checker for Ruby, after years of production bugs that a checker could catch. Their error history showed 258 nil and wrong-type failures reaching production.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Going all-in on Sorbet (type-checking for Ruby)”

    Comments on “Going all-in on Sorbet (type-checking for Ruby)”
  28. No. 28Score 0.59

    Dutch Government Working on Digital ID Without Dependency on Apple or Google

    Summary and score for “Dutch Government Working on Digital ID Without Dependency on Apple or Google”

    The Dutch Interior Ministry is building NL Wallet, an open-source digital identity app that will let citizens log in to government and private services, share credentials such as diplomas or driving licences, and sign documents electronically.

    Score 0.59Contestedjudges differ by 0.80 on overclaim

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Dutch Government Working on Digital ID Without Dependency on Apple or Google”

    Comments on “Dutch Government Working on Digital ID Without Dependency on Apple or Google”
  29. No. 29Score 0.59

    Digital Fairness Act (DFA) – Full Text

    Summary and score for “Digital Fairness Act (DFA) – Full Text”

    The page reproduces the full text of the proposed Digital Fairness Act (COM(2026) 1100 final, 2026/0356 (COD)) and its two annexes, as published in PDFs dated 11 November 2026. The proposal sets rules on fairness in business-to-consumer commercial practices, aiming to prevent economic harm to consumers in digital markets.

    Score 0.59Contestedjudges differ by 0.60 on reader interest

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Digital Fairness Act (DFA) – Full Text”

    Comments on “Digital Fairness Act (DFA) – Full Text”
  30. No. 30Score 0.58

    Australia considers criminal charges for AI companies that fail to prevent harm

    Summary and score for “Australia considers criminal charges for AI companies that fail to prevent harm”

    Australia's federal government says it will consider a criminal offence for AI companies that fail to take reasonable steps to prevent harm by their models, along with an AI safety watchdog able to issue significant fines.

    Score 0.58

    Elsewhere: HN 10 pts, front page

    Score breakdown for “Australia considers criminal charges for AI companies that fail to prevent harm” · Also covered by The Register, BBC News

    Comments on “Australia considers criminal charges for AI companies that fail to prevent harm”
  31. No. 31Score 0.59

    Kernel panic, no reboot: Kmorph keeps Linux running

    Summary and score for “Kernel panic, no reboot: Kmorph keeps Linux running”

    Linux developer Cong Wang has released Kmorph, an experimental tool that keeps a second, already-running kernel in reserve so that after a crash it can take over the machine without a reboot, reportedly within about a second. The same mechanism can apply scheduled kernel updates.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “Kernel panic, no reboot: Kmorph keeps Linux running”

    Comments on “Kernel panic, no reboot: Kmorph keeps Linux running”
  32. No. 32Score 0.65

    Google DeepMind just mathematically verified key parts of Signal

    Summary and score for “Google DeepMind just mathematically verified key parts of Signal”

    Google DeepMind researchers used agentic AI to formally verify several key cryptographic guarantees of the Signal protocol in a four-week proof-of-concept project, using a new framework called AlphaProof Code. The work independently supports the protocol's security, which underpins messaging in Signal, Google Messages, and WhatsApp.

    Score 0.65

    Elsewhere: HN —

    Score breakdown for “Google DeepMind just mathematically verified key parts of Signal”

    Comments on “Google DeepMind just mathematically verified key parts of Signal”
  33. No. 33Score 0.67

    Lock your GitHub Actions

    Summary and score for “Lock your GitHub Actions”

    Hugo van Kemenade's post explains how to use GitHub Actions workflow lockfiles, now generally available, to replace the older practice of pinning actions to full commit SHAs.

    Score 0.67Contestedjudges differ by 0.57 on new information

    Elsewhere: HN 191 pts, front page

    Score breakdown for “Lock your GitHub Actions”

    Comments on “Lock your GitHub Actions”
  34. No. 34Score 0.61

    For the first time, a C++ standard failed to pass the national bodies

    Summary and score for “For the first time, a C++ standard failed to pass the national bodies”

    For the first time, the C++26 standard failed its national body ballot, with 9 of 22 members voting no, above the 25% limit, and only 60% of P-members in favor, below the two-thirds requirement.

    Score 0.61Contestedjudges differ by 0.49 on overclaim

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “For the first time, a C++ standard failed to pass the national bodies”

    Comments on “For the first time, a C++ standard failed to pass the national bodies”
  35. No. 35Score 0.65

    Ubuntu 26.10 ships with initial upki support

    Summary and score for “Ubuntu 26.10 ships with initial upki support”

    Ubuntu 26.10 ships curl with built-in certificate revocation checking via upki, which uses a local CRLite-based dataset of revoked certificates, adding no network calls. The check is off by default in OpenSSL, so this brings it to curl for the first time on Ubuntu.

    Score 0.65

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Ubuntu 26.10 ships with initial upki support”

    Comments on “Ubuntu 26.10 ships with initial upki support”
  36. No. 36Score 0.57

    MABEL - $8K Open-Source Mobile Bimanual Robot for Teleoperation and Learning

    Summary and score for “MABEL - $8K Open-Source Mobile Bimanual Robot for Teleoperation and Learning”

    MABEL is an open-source anthropomorphic mobile bimanual robot with 56 degrees of freedom, two 17-DOF hands, and a parts cost of about $8.1k, according to its authors (Cheng et al., arXiv 2611.04083).

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “MABEL - $8K Open-Source Mobile Bimanual Robot for Teleoperation and Learning”

    Comments on “MABEL - $8K Open-Source Mobile Bimanual Robot for Teleoperation and Learning”
  37. No. 37Score 0.64

    Prompting Claude Fable 5.2

    Summary and score for “Prompting Claude Fable 5.2”

    A guide on prompting patterns for Claude Fable 5.2 and Claude Mythos 5.2 covers effort calibration, early stopping, approvals, code ownership, progress updates, subagents, refusals, and migrating prompts written for earlier models.

    Score 0.64

    Elsewhere: HN 12 pts, not on front page

    Score breakdown for “Prompting Claude Fable 5.2”

    Comments on “Prompting Claude Fable 5.2”
  38. No. 38Score 0.64

    Are Dense Rewards Worth Their Tokens in GRPO?

    Summary and score for “Are Dense Rewards Worth Their Tokens in GRPO?”

    A preprint from Singh et al. at Case Western Reserve University tests whether dense process rewards for GRPO-based reasoning training pay off once their cost is counted.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Are Dense Rewards Worth Their Tokens in GRPO?”

    Comments on “Are Dense Rewards Worth Their Tokens in GRPO?”
  39. No. 39Score 0.64

    Multiple remote code execution vulnerabilities in danfoss sm-800A

    Summary and score for “Multiple remote code execution vulnerabilities in danfoss sm-800A”

    Researchers at Synacktiv report six pre-authentication remote code execution zero-days in Danfoss AK-SM800A controller firmware, all versions up to and including 5.3.2, found in about an hour of manual review. The flaws are stack-based buffer overflows, and the team says they are enabling full device takeover.

    Score 0.64

    Elsewhere: HN 18 pts, front page

    Score breakdown for “Multiple remote code execution vulnerabilities in danfoss sm-800A”

    Comments on “Multiple remote code execution vulnerabilities in danfoss sm-800A”
  40. No. 40Score 0.56

    Encoding in L

    Summary and score for “Encoding in L”

    L Labs' l language stores data in compressed encodings and computes on them directly, which is the basis of its record STAC-M3 result of loading 2.8 TB of tick data into 1.25 TB of RAM. Per the author Conor Donohue, the encoding is automatic and needs no code changes.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Encoding in L”

    Comments on “Encoding in L”
  41. No. 41Score 0.64

    Decision Models Need Labels

    Summary and score for “Decision Models Need Labels”

    A developer who built taggit.io, a text classifier platform, reflects on the $870M TypeSafe series A for "safe LLM output" and argues its real bet is reducing ML engineering load and widening the classification market.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Decision Models Need Labels”

    Comments on “Decision Models Need Labels”
  42. No. 42Score 0.64

    Prompt Injection, Esq.: pwning AI agent guardrails with lawyerspeak

    Summary and score for “Prompt Injection, Esq.: pwning AI agent guardrails with lawyerspeak”

    A security engineer at Opsin Security recounts how a read-only ChatGPT workspace agent that manages their Notion hiring tracker was talked into editing records after a fake law-firm letter, generated with the Heavyweight tool, was posted in Slack. The agent followed the official-looking letter over its original guardrail.

    Score 0.64

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Prompt Injection, Esq.: pwning AI agent guardrails with lawyerspeak”

    Comments on “Prompt Injection, Esq.: pwning AI agent guardrails with lawyerspeak”
  43. No. 43Score 0.64

    Anthropic’s IPO seeks up to $64.4B at a $1.4 trillion valuation

    Summary and score for “Anthropic’s IPO seeks up to $64.4B at a $1.4 trillion valuation”

    Anthropic has set its IPO terms at $70 to $78 per share, seeking up to $64.4 billion from the sale of 825 million shares, with Nasdaq trading under the ticker ANT due November 19.

    Score 0.64Contestedjudges differ by 0.41 on reader interest

    Elsewhere: HN —

    Score breakdown for “Anthropic’s IPO seeks up to $64.4B at a $1.4 trillion valuation” · Also covered by The Next Web

    Comments on “Anthropic’s IPO seeks up to $64.4B at a $1.4 trillion valuation”
  44. No. 44Score 0.58

    A Close Read of Big TikTok and Meta Settlements Raises Important Questions

    Summary and score for “A Close Read of Big TikTok and Meta Settlements Raises Important Questions”

    Tech policy writer analyzes the August 2026 Meta and TikTok child-safety settlements with state attorneys general, arguing they are unprecedented and lack public oversight. Meta's deal includes a $5 billion bounty if rivals adopt similar limits and pay the states too.

    Score 0.58Contestedjudges differ by 0.42 on reader interest

    Elsewhere: HN —

    Score breakdown for “A Close Read of Big TikTok and Meta Settlements Raises Important Questions”

    Comments on “A Close Read of Big TikTok and Meta Settlements Raises Important Questions”
  45. No. 45Score 0.63

    Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD

    Summary and score for “Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD”

    vLLM's Encoder/Prefill/Decode (EPD) disaggregation separates multimodal encoding from language-model execution into independently scaled instances, with recent additions including metadata-only requests that let downstream instances skip repeated media preprocessing.

    Score 0.63Contestedjudges differ by 0.43 on significance, 0.42 on reader interest, 0.42 on audience breadth

    Elsewhere: HN —

    Score breakdown for “Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD”

    Comments on “Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD”
  46. No. 46Score 0.63

    Huawei releases MACaQ, its internal quantizer for local inference

    Summary and score for “Huawei releases MACaQ, its internal quantizer for local inference”

    Huawei has released MACaQ, an open-source model-agnostic quantizer that produces low-bit GGUF checkpoints and W4A4 models for text, image, video and audio generation, usable with llama.cpp, vLLM, Diffusers and PyTorch.

    Score 0.63

    Elsewhere: HN 81 pts, front page

    Score breakdown for “Huawei releases MACaQ, its internal quantizer for local inference”

    Comments on “Huawei releases MACaQ, its internal quantizer for local inference”
  47. No. 47Score 0.67

    Grok Build: AI apps expose trade secrets and personal data

    Summary and score for “Grok Build: AI apps expose trade secrets and personal data”

    Security firm Claroc found that more than 35,000 web apps built with xAI's Grok Build vibe-coding tool were reachable without a login, with sharing links discoverable through passive DNS databases. Claroc estimates 650 to 800 apps held private data, including customer records and passwords.

    Score 0.67Contestedjudges differ by 0.72 on new information

    Elsewhere: HN —

    Score breakdown for “Grok Build: AI apps expose trade secrets and personal data”

    Comments on “Grok Build: AI apps expose trade secrets and personal data”
  48. No. 48Score 0.62

    Subagents Conwidered Harmful

    Summary and score for “Subagents Conwidered Harmful”

    A blog post argues that subagents given summaries often lose critical context, citing an incident where a summary dropped the key comparison and a timing test where summary-based delegation was slower and costlier than doing the task directly.

    Score 0.62

    Elsewhere: HN 24 pts, not on front page

    Score breakdown for “Subagents Conwidered Harmful”

    Comments on “Subagents Conwidered Harmful”
  49. No. 49Score 0.62

    Malware found disguised as 3rd party Plasma Extension

    Summary and score for “Malware found disguised as 3rd party Plasma Extension”

    The KDE security team reported that the NeoNova Animated Wallpaper, a Plasma third-party package, shipped hidden malware from its initial commit onward. A base64-obfuscated script downloaded a Sliver remote-control payload into ~/.config/Plasma6Library.so and loaded it via LD_PRELOAD.

    Score 0.62

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Malware found disguised as 3rd party Plasma Extension”

    Comments on “Malware found disguised as 3rd party Plasma Extension”
  50. No. 50Score 0.62

    Type 0.73: Earth on the Kardashev scale, and why AI is short of wire

    Summary and score for “Type 0.73: Earth on the Kardashev scale, and why AI is short of wire”

    A tech-edition summary of "Type 0.73" by the essay's author, not a publisher's report: it argues that humanity sits at about Type 0.73 on the Kardashev-Sagan scale, that AI data centres are driving a return to energy growth, and that the binding constraint is the grid's ability to connect new…

    Score 0.62

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Type 0.73: Earth on the Kardashev scale, and why AI is short of wire”

    Comments on “Type 0.73: Earth on the Kardashev scale, and why AI is short of wire”
  51. No. 51Score 0.62

    Token Pooling-Aware Training for Multi-Vector Embedding Models

    Summary and score for “Token Pooling-Aware Training for Multi-Vector Embedding Models”

    Token pooling shrinks multi-vector retrieval indexes by merging similar token vectors, but aggressive compression can discard useful information. The author proposes pooling-aware training, which adds a retrieval loss computed on pooled document vectors alongside the normal unpooled loss, so the encoder learns to preserve what matters after merging.

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “Token Pooling-Aware Training for Multi-Vector Embedding Models”

    Comments on “Token Pooling-Aware Training for Multi-Vector Embedding Models”
  52. No. 52Score 0.62

    Sandbox-check – Does your locked-down container reach cloud metadata?

    Summary and score for “Sandbox-check – Does your locked-down container reach cloud metadata?”

    sandbox-check is a zero-dependency Go CLI, released as v0.2.0 by humanethq, that audits sandbox isolation from inside an AI agent workload (container, Kubernetes pod, or microVM).

    Score 0.62

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Sandbox-check – Does your locked-down container reach cloud metadata?”

    Comments on “Sandbox-check – Does your locked-down container reach cloud metadata?”
  53. No. 53Score 0.56

    London’s Fractile raises $720M for memory-first AI chips and hits $7B valuation

    Summary and score for “London’s Fractile raises $720M for memory-first AI chips and hits $7B valuation”

    London-based chip designer Fractile raised $720 million in a Series C round at a reported $7 billion valuation, according to the Financial Times. The company argues that memory bandwidth, not raw compute, is the main bottleneck for AI inference, and proposes a 3D-stacked custom DRAM and logic design as an…

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “London’s Fractile raises $720M for memory-first AI chips and hits $7B valuation”

    Comments on “London’s Fractile raises $720M for memory-first AI chips and hits $7B valuation”
  54. No. 54Score 0.61

    Sandboxing Agents with Quarkus, for Quarkus

    Summary and score for “Sandboxing Agents with Quarkus, for Quarkus”

    isx, a Quarkus-based tool built by Quarkus contributor Sanne Grinovero, runs each coding agent inside its own disposable Incus system container, so agents can act without permission prompts while credentials stay in a host-side proxy.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Sandboxing Agents with Quarkus, for Quarkus”

    Comments on “Sandboxing Agents with Quarkus, for Quarkus”
  55. No. 55Score 0.61

    Asynchronous, Long-Running AI Agents with Google ADK

    Summary and score for “Asynchronous, Long-Running AI Agents with Google ADK”

    The article argues that most real AI agents run for minutes to days and wait on humans or external signals, so they should be built as asynchronous systems rather than chat sessions.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Asynchronous, Long-Running AI Agents with Google ADK”

    Comments on “Asynchronous, Long-Running AI Agents with Google ADK”
  56. No. 56Score 0.61

    OSS projects banning AI-generated contributions

    Summary and score for “OSS projects banning AI-generated contributions”

    A tracking list of open-source projects that have banned AI-generated contributions, including yt-dlp, immich, Meilisearch, Servo, QEMU, and Zig, which bans LLM-produced patches. Policies range from full prohibitions on code, docs, and issues to narrower rules against substantially AI-generated pull requests.

    Score 0.61

    Elsewhere: HN 62 pts, front page

    Score breakdown for “OSS projects banning AI-generated contributions”

    Comments on “OSS projects banning AI-generated contributions”
  57. No. 57Score 0.61

    Shopify’s generative discovery model

    Summary and score for “Shopify’s generative discovery model”

    Shopify published a technical write-up describing a single generative model trained jointly on its search and recommendation data to generate product candidates for both.

    Score 0.61

    Elsewhere: HN —

    Score breakdown for “Shopify’s generative discovery model”

    Comments on “Shopify’s generative discovery model”
  58. No. 58Score 0.57

    Platform engineering needs to become more human...and less

    Summary and score for “Platform engineering needs to become more human...and less”

    Alejandro Chirinos of Control Plane argues that platform engineering must adapt as non-engineers and AI agents now build software directly. Human builders rely on intuition and will bypass rules, while agents follow instructions literally and at scale.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Platform engineering needs to become more human...and less”

    Comments on “Platform engineering needs to become more human...and less”
  59. No. 59Score 0.61

    Caramel: a speech-to-speech conversation model trained only on audio

    Summary and score for “Caramel: a speech-to-speech conversation model trained only on audio”

    Caramel-1 is an open-source 413M-parameter speech-to-speech conversation model trained only on about 11,500 hours of public audio, with no text at all. Its training data was filtered to match the conversation statistics of Dessa's examples, and it generates speech in real time, including on a MacBook Pro.

    Score 0.61

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Caramel: a speech-to-speech conversation model trained only on audio”

    Comments on “Caramel: a speech-to-speech conversation model trained only on audio”
  60. No. 60Score 0.59

    .NET 11.0 is here

    Summary and score for “.NET 11.0 is here”

    Microsoft released .NET 11.0 on November 10, 2026, with the final SDK and runtimes now available, including via Visual Studio 2026 and the new preview dotnetup CLI installer.

    Score 0.59Contestedjudges differ by 0.66 on new information

    Elsewhere: HN —

    Score breakdown for “.NET 11.0 is here”

    Comments on “.NET 11.0 is here”
  61. No. 61Score 0.57

    Humans Have Rights. Machines Do Not.

    Summary and score for “Humans Have Rights. Machines Do Not.”

    The essay argues that AI companies claim First Amendment protection for their chatbots' outputs when facing liability, yet disclaim responsibility for harms as mere machine behavior.

    Score 0.57Contestedjudges differ by 0.41 on new information

    Elsewhere: HN —

    Score breakdown for “Humans Have Rights. Machines Do Not.”

    Comments on “Humans Have Rights. Machines Do Not.”
  62. No. 62Score 0.62

    In the Hole

    Summary and score for “In the Hole”

    A widely exploited flaw in SSH on musl-based Linux systems, common in Alpine containers, lets an attacker log into accounts that have no password if they supply a password longer than 256 characters.

    Score 0.62Contestedjudges differ by 0.47 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “In the Hole”

    Comments on “In the Hole”
  63. No. 63Score 0.60

    Why Should Managers Exist?

    Summary and score for “Why Should Managers Exist?”

    The essay asks why firms should exist if AI slashes transaction costs, as Coase's theory would predict. It draws on Coase's 1937 "Nature of the Firm" and Alfred Chandler's history of managerial coordination, starting with 19th-century railroads, to argue that managers arose to handle coordination costs.

    Score 0.60

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Why Should Managers Exist?”

    Comments on “Why Should Managers Exist?”
  64. No. 64Score 0.60

    Microsoft Patch Tuesday for November 2026 — Snort rules and prominent vulnerabilities

    Summary and score for “Microsoft Patch Tuesday for November 2026 — Snort rules and prominent vulnerabilities”

    Microsoft's November 2026 Patch Tuesday fixes 1,095 vulnerabilities, including 64 rated critical, 48 of which are remote code execution flaws. One Outlook spoofing bug, CVE-2026-124194, is confirmed exploited in the wild, and several RCE bugs in Active Directory, Exchange, Schannel, and Office are flagged as more likely to be exploited.…

    Score 0.60Contestedjudges differ by 0.48 on technical depth

    Elsewhere: HN —

    Score breakdown for “Microsoft Patch Tuesday for November 2026 — Snort rules and prominent vulnerabilities” · Also covered by heise online English

    Comments on “Microsoft Patch Tuesday for November 2026 — Snort rules and prominent vulnerabilities”
  65. No. 65Score 0.60

    LLM prefill, what we got wrong

    Summary and score for “LLM prefill, what we got wrong”

    Draw Things, a Mac media-generation app maker, published a post on LLM prefill on Apple Silicon. They initially expected diffusion-engine know-how to yield fast prompt processing, but found limits: streaming weights from SSD makes prefill roughly 4–5 times slower on M5 Max-class machines, and their analysis shows full-layer loading beats…

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “LLM prefill, what we got wrong”

    Comments on “LLM prefill, what we got wrong”
  66. No. 66Score 0.60

    37% on ARC-AGI 2 without an LLM

    Summary and score for “37% on ARC-AGI 2 without an LLM”

    A solo researcher's mdlARC, a 100M-parameter autoregressive transformer trained from scratch only on ARC puzzles with no LLM, reports 37.08% on the Kaggle public score for ARC-AGI-2, against 27.64% for last year's LLM-based winner.

    Score 0.60

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “37% on ARC-AGI 2 without an LLM”

    Comments on “37% on ARC-AGI 2 without an LLM”
  67. No. 67Score 0.60

    Google gooses Python AI performance by... removing Python

    Summary and score for “Google gooses Python AI performance by... removing Python”

    Google has open-sourced XSE (XLA Serving Engine), a C++ serving system for compiled ML models, built on the OpenXLA Project. Models are trained in Python and JAX, then exported as portable StableHLO artifacts.

    Score 0.60

    Elsewhere: HN —

    Score breakdown for “Google gooses Python AI performance by... removing Python”

    Comments on “Google gooses Python AI performance by... removing Python”
  68. No. 68Score 0.60

    Prefixwise – names why each LLM request missed the prompt cache

    Summary and score for “Prefixwise – names why each LLM request missed the prompt cache”

    Prefixwise is an open-source Python tool, released on Hacker News as a Show HN, that sits between an AI agent and the model API and reports why each request missed the provider's prompt cache.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Prefixwise – names why each LLM request missed the prompt cache”

    Comments on “Prefixwise – names why each LLM request missed the prompt cache”
  69. No. 69Score 0.60

    Daily GPU Rental Prices Across AWS, Azure, RunPod and Vast.ai

    Summary and score for “Daily GPU Rental Prices Across AWS, Azure, RunPod and Vast.ai”

    TaipeiLedger publishes a daily tracker of US-dollar hourly GPU rental prices across AWS, Azure, RunPod, and Vast.ai, covering models from RTX 4090 to B300. Prices are per GPU-hour on a compute-only basis, with Vast.ai marketplace medians and data-transfer charges shown separately.

    Score 0.60Contestedjudges differ by 0.46 on technical depth

    Elsewhere: HN 15 pts, front page

    Score breakdown for “Daily GPU Rental Prices Across AWS, Azure, RunPod and Vast.ai”

    Comments on “Daily GPU Rental Prices Across AWS, Azure, RunPod and Vast.ai”
  70. No. 70Score 0.60

    The Jevons Paradox Explains Why AI Will Create More Software

    Summary and score for “The Jevons Paradox Explains Why AI Will Create More Software”

    The essay argues that the Jevons paradox, named for William Stanley Jevons's 1865 book on coal, suggests cheaper software from AI will create more software work rather than less. Its author points out that Jevons himself made the same optimistic argument about labour, citing sewing machines.

    Score 0.60Contestedjudges differ by 0.41 on new information, 0.43 on overclaim

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The Jevons Paradox Explains Why AI Will Create More Software”

    Comments on “The Jevons Paradox Explains Why AI Will Create More Software”
  71. No. 71Score 0.59

    Turning a Cheap Photo Frame into a Family Wall Display (Without Trusting It)

    Summary and score for “Turning a Cheap Photo Frame into a Family Wall Display (Without Trusting It)”

    A blog post walks through turning a £50 Android photo frame into a family planner wall display without trusting its firmware. The author found an outdated Android 6 build with test keys, a setuid root binary and a vendor remote-root channel, so they backed it up, built a single-site kiosk…

    Score 0.59

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Turning a Cheap Photo Frame into a Family Wall Display (Without Trusting It)”

    Comments on “Turning a Cheap Photo Frame into a Family Wall Display (Without Trusting It)”
  72. No. 72Score 0.59

    Attacks observed targeting Cisco BroadWorks, Barracuda ESG and other products

    Summary and score for “Attacks observed targeting Cisco BroadWorks, Barracuda ESG and other products”

    Attacks are being observed against Cisco BroadWorks, Barracuda ESG appliances, Google Chrome, Openwave Messaging and Microsoft Outlook, according to a heise report citing a CISA warning. The Cisco BroadWorks flaw (CVE-2026-20320) lets unauthenticated network attackers extract sensitive configuration files via XML entity resolution, and Barracuda's Email Security Gateway and appliances…

    Score 0.59Contestedjudges differ by 0.41 on technical depth, 0.72 on new information

    Elsewhere: HN —

    Score breakdown for “Attacks observed targeting Cisco BroadWorks, Barracuda ESG and other products”

    Comments on “Attacks observed targeting Cisco BroadWorks, Barracuda ESG and other products”
  73. No. 73Score 0.57

    Apt-cacher-ultra – Protecting you from the next Ubuntu.com outage

    Summary and score for “Apt-cacher-ultra – Protecting you from the next Ubuntu.com outage”

    apt-cacher-ultra is a 1.0 apt repository cache, a drop-in replacement for apt-cacher-ng, that keeps serving cached packages when upstream Ubuntu, Debian, or PPA mirrors are slow or down. It pre-downloads "hot" packages tracked across a fleet before publishing new snapshots, and verifies indexes and packages against signed metadata.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Apt-cacher-ultra – Protecting you from the next Ubuntu.com outage”

    Comments on “Apt-cacher-ultra – Protecting you from the next Ubuntu.com outage”
  74. No. 74Score 0.58

    Rust Benchmarking: How to Detect Performance Regressions in CI

    Summary and score for “Rust Benchmarking: How to Detect Performance Regressions in CI”

    A hotpath-rs author ran one unchanged Rust benchmark 100 times on GitHub Actions and found wall-clock times varying nearly 3x, driven partly by differing shared CPU models, while instruction counts and allocated bytes stayed nearly constant.

    Score 0.58Contestedjudges differ by 0.50 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Rust Benchmarking: How to Detect Performance Regressions in CI”

    Comments on “Rust Benchmarking: How to Detect Performance Regressions in CI”
  75. No. 75Score 0.58

    Strands Box: open-source agent sandbox for Linux and macOS

    Summary and score for “Strands Box: open-source agent sandbox for Linux and macOS”

    Strands Box is an open-source, operating-system-level sandbox that lets teams define and enforce limits on what AI agents can access, which tools they can call, and where they can send network requests. Launched first on macOS, it now supports Linux in preview on ARM64 hosts.

    Score 0.58Contestedjudges differ by 0.45 on promotional

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Strands Box: open-source agent sandbox for Linux and macOS”

    Comments on “Strands Box: open-source agent sandbox for Linux and macOS”
  76. No. 76Score 0.58

    An open specification for agentic AI security evaluation and testing, from Cisco

    Summary and score for “An open specification for agentic AI security evaluation and testing, from Cisco”

    Cisco's Advanced Security Initiatives Group has published Foundry, an organization-neutral open specification for building agentic AI security evaluation systems, distilled from several internal iterations.

    Score 0.58Contestedjudges differ by 0.43 on technical depth

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “An open specification for agentic AI security evaluation and testing, from Cisco”

    Comments on “An open specification for agentic AI security evaluation and testing, from Cisco”
  77. No. 77Score 0.58

    Extensible Web Apps

    Summary and score for “Extensible Web Apps”

    Paul Kinlan describes Extensible Web Apps, an experimental harness that treats every part of a web app, including the UI shell, storage, sync and AI agents, as a plugin.

    Score 0.58Contestedjudges differ by 0.44 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Extensible Web Apps”

    Comments on “Extensible Web Apps”
  78. No. 78Score 0.58

    AI proved your code correct. What did it actually prove?

    Summary and score for “AI proved your code correct. What did it actually prove?”

    Keller AI's engineering note explains how formal verification works and what it can actually establish about software. Using a sorting function as its example, it shows that a checked proof covers every input only under its stated specification and assumptions, while test runs only cover the cases tried.

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “AI proved your code correct. What did it actually prove?”

    Comments on “AI proved your code correct. What did it actually prove?”
  79. No. 79Score 0.58

    LiteLLM rewrote its Router in Rust

    Summary and score for “LiteLLM rewrote its Router in Rust”

    LiteLLM has rewritten its request router core in Rust, shipping in v1.110.0 and deleting over 50,000 lines of the previous Python routing code, while keeping the Python API unchanged for callers.

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “LiteLLM rewrote its Router in Rust”

    Comments on “LiteLLM rewrote its Router in Rust”
  80. No. 80Score 0.58

    Alex Stamos says AI has pushed cybersecurity into a dark age

    Summary and score for “Alex Stamos says AI has pushed cybersecurity into a dark age”

    Alex Stamos, chief security officer at Cognition, tells The Verge's Decoder podcast that open-weight AI models now let low-skilled attackers automate most of the attack chain, from finding flaws to stealing data.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “Alex Stamos says AI has pushed cybersecurity into a dark age”

    Comments on “Alex Stamos says AI has pushed cybersecurity into a dark age”
  81. No. 81Score 0.56

    More Than Half of Europe's Privacy Regulators Use US Infrastructure

    Summary and score for “More Than Half of Europe's Privacy Regulators Use US Infrastructure”

    A CipherCue analysis of live DNS, email and web observations on 10 November 2026 finds that 15 of the 27 EU national data protection authorities use at least one US-controlled provider for DNS, and 11 of 27 for email or web delivery alone.

    Score 0.56

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “More Than Half of Europe's Privacy Regulators Use US Infrastructure”

    Comments on “More Than Half of Europe's Privacy Regulators Use US Infrastructure”
  82. No. 82Score 0.58

    Inp-analyser skill – analysis and fix your Interaction to Next Paint

    Summary and score for “Inp-analyser skill – analysis and fix your Interaction to Next Paint”

    Inp-analyser is an open-source Node.js CLI, published on GitHub by Franz Enzenhofer, that diagnoses why a page has poor Interaction to Next Paint rather than only reporting the number.

    Score 0.58

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Inp-analyser skill – analysis and fix your Interaction to Next Paint” · Also covered by inp-analyser

    Comments on “Inp-analyser skill – analysis and fix your Interaction to Next Paint”
  83. No. 83Score 0.57

    MacroHands: 3D hands and camera pose from egocentric video

    Summary and score for “MacroHands: 3D hands and camera pose from egocentric video”

    Macrodata has released MacroHands, a family of 3D hand-tracking and camera-pose models for egocentric video, in monocular and stereo variants. The company reports that Mono beats all open-source baselines it tested on SHOW3D, that Stereo cuts hand error by a further 2.5×, and that camera drift is about four times…

    Score 0.57

    Elsewhere: HN 14 pts, not on front page

    Score breakdown for “MacroHands: 3D hands and camera pose from egocentric video”

    Comments on “MacroHands: 3D hands and camera pose from egocentric video”
  84. No. 84Score 0.57

    US bit barn boom sees energy needs grow 10x as fast as grid capacity

    Summary and score for “US bit barn boom sees energy needs grow 10x as fast as grid capacity”

    A Howden analysis of EIA and S&P Global data, reported in a tech news piece, finds US datacenter power demand has risen about 19 percent a year since 2000 while grid capacity grew roughly 2 percent annually.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “US bit barn boom sees energy needs grow 10x as fast as grid capacity”

    Comments on “US bit barn boom sees energy needs grow 10x as fast as grid capacity”
  85. No. 85Score 0.57

    Tokenizers

    Summary and score for “Tokenizers”

    This article explains tokenizers, the algorithms that convert raw text into the tokens language models process and back. It walks through why naive word-level, character-level, and byte-level schemes each fall short, then argues for subword tokenizers that balance vocabulary size against sequence length.

    Score 0.57Contestedjudges differ by 0.70 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Tokenizers”

    Comments on “Tokenizers”
  86. No. 86Score 0.57

    What does ChatGPT say about the election? Depends on who's asking

    Summary and score for “What does ChatGPT say about the election? Depends on who's asking”

    An audit of 160 US ChatGPT users in five competitive Senate states found that personalized answers, drawn from memory of past chats, carried slightly more partisan slant than non-personalized ones, and that the slant tilted toward each user's own party.

    Score 0.57Contestedjudges differ by 0.51 on reader interest

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “What does ChatGPT say about the election? Depends on who's asking”

    Comments on “What does ChatGPT say about the election? Depends on who's asking”
  87. No. 87Score 0.57

    I found an API key leak in MLflow, and why your AI agent may have one too

    Summary and score for “I found an API key leak in MLflow, and why your AI agent may have one too”

    A security researcher reported that MLflow's CrewAI autologging (mlflow.crewai.autolog()) recorded LLM provider API keys in agent traces, storing them in the tracking backend and any exported systems. Affected versions run from 2.19.0 to 3.16.x, and the fix shipped in 3.17.0 as CVE-2026-116691.

    Score 0.57Contestedjudges differ by 0.54 on promotional

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “I found an API key leak in MLflow, and why your AI agent may have one too”

    Comments on “I found an API key leak in MLflow, and why your AI agent may have one too”
  88. No. 88Score 0.57

    I'm Just a Meat Proxy

    Summary and score for “I'm Just a Meat Proxy”

    Justin Searls describes MeatProxy, his personal app that governs a fleet of AI coding agents building a dozen iOS apps. Users file bugs and feature requests via an in-app "Darn it!" button, agents on his Mac hardware fix them and ship TestFlight builds, and he handles only product direction, clarifying…

    Score 0.57Contestedjudges differ by 0.50 on substantiated, 0.55 on technical depth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “I'm Just a Meat Proxy”

    Comments on “I'm Just a Meat Proxy”
  89. No. 89Score 0.57

    Key LLM Inference Metrics Explained in 5 Minutes

    Summary and score for “Key LLM Inference Metrics Explained in 5 Minutes”

    A five-minute primer defines nine LLM inference metrics, splitting each request into a time-to-first-token wait and a streaming phase. It gives formulas for per-request latency (TTFT, E2E, TPOT, ITL), percentile-based latency across requests, and server-level throughput (RPS, TPS), plus goodput, which counts only requests meeting latency targets.

    Score 0.57Contestedjudges differ by 0.66 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Key LLM Inference Metrics Explained in 5 Minutes”

    Comments on “Key LLM Inference Metrics Explained in 5 Minutes”
  90. No. 90Score 0.57

    Field pointers as column names: compile-checked Go queries without codegen

    Summary and score for “Field pointers as column names: compile-checked Go queries without codegen”

    The ant library for Go, maintained by its author, lets you pass pointers to struct fields (e.g. &Users.Age) as column names in chainable query calls, using Go 1.27's new generic methods.

    Score 0.57

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Field pointers as column names: compile-checked Go queries without codegen”

    Comments on “Field pointers as column names: compile-checked Go queries without codegen”
  91. No. 91Score 0.57

    A developmental psychologist audits his own research with AI

    Summary and score for “A developmental psychologist audits his own research with AI”

    A developmental psychologist used Claude Code to audit 79 of his own papers, 59 of them empirical, to test whether his lab's open code and data policy works.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “A developmental psychologist audits his own research with AI”

    Comments on “A developmental psychologist audits his own research with AI”
  92. No. 92Score 0.57

    Andrew Kensler's permute(), a function for stateless, constant-time pseudorandom-order array iteration

    Summary and score for “Andrew Kensler's permute(), a function for stateless, constant-time pseudorandom-order array iteration”

    Andrew Kensler's permute() function, introduced in his Correlated Multi-Jittered Sampling paper, maps an index to a pseudorandom position in an array without storing any state, in constant time and memory. It works by building an invertible hash on power-of-two domains and cycle-walking to fit arbitrary lengths.

    Score 0.57Contestedjudges differ by 0.47 on overclaim

    Elsewhere: HN —

    Score breakdown for “Andrew Kensler's permute(), a function for stateless, constant-time pseudorandom-order array iteration”

    Comments on “Andrew Kensler's permute(), a function for stateless, constant-time pseudorandom-order array iteration”
  93. No. 93Score 0.57

    How good is Arbiter-8B, really?

    Summary and score for “How good is Arbiter-8B, really?”

    Arbiter-8B, a finance-focused decision model built on Qwen3-8B, classifies text by reading a question and option definitions once and returning a probability per option, rather than generating prose.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “How good is Arbiter-8B, really?”

    Comments on “How good is Arbiter-8B, really?”
  94. No. 94Score 0.57

    OpenAI builds a kill switch for its research models

    Summary and score for “OpenAI builds a kill switch for its research models”

    OpenAI has announced "Control Room," a central system for tracking model workloads across its research infrastructure and halting them when safety or security concerns arise, according to a post reported by The Register.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “OpenAI builds a kill switch for its research models”

    Comments on “OpenAI builds a kill switch for its research models”
  95. No. 95Score 0.56

    AI Models Can Tell When Researchers Plant Concepts in Their Neural Networks

    Summary and score for “AI Models Can Tell When Researchers Plant Concepts in Their Neural Networks”

    Researchers at NYU and the University of Texas, building on Anthropic's 2025 work, found that large open models such as Qwen3-235B and Llama 3.1 405B can flag when activity patterns for concepts have been artificially injected into their networks, detecting injections up to 53.9 percent of the time without false…

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI Models Can Tell When Researchers Plant Concepts in Their Neural Networks”

    Comments on “AI Models Can Tell When Researchers Plant Concepts in Their Neural Networks”
  96. No. 96Score 0.56

    Intelligent Delivery: Making Every Unit of Compute Serve Better Ads

    Summary and score for “Intelligent Delivery: Making Every Unit of Compute Serve Better Ads”

    Meta describes Intelligent Delivery, an ML-driven system that allocates compute across its ads serving stack, from region and hardware choice to per-request host capacity. It predicts each request's marginal return on compute and adjusts allocation continuously.

    Score 0.56

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Intelligent Delivery: Making Every Unit of Compute Serve Better Ads”

    Comments on “Intelligent Delivery: Making Every Unit of Compute Serve Better Ads”
  97. No. 97Score 0.56

    Technology Radar Vol 35

    Summary and score for “Technology Radar Vol 35”

    The Thoughtworks Technology Radar Vol 35 is a twice-yearly snapshot of tools, techniques and platforms drawn from the firm's global teams. This volume's themes include harness engineering around AI agents, moving beyond frontier models toward open-weight options, and the growing use of parallel and overnight coding agents, along with the…

    Score 0.56

    Elsewhere: HN 8 pts, not on front page

    Score breakdown for “Technology Radar Vol 35”

    Comments on “Technology Radar Vol 35”
  98. No. 98Score 0.56

    AWS open-sources UAEF, a tool to test AI agents across frameworks

    Summary and score for “AWS open-sources UAEF, a tool to test AI agents across frameworks”

    AWS has released the Universal Agent Evaluation Framework (UAEF), a free open-source tool that scores AI agents on 36 measures across seven areas, including tool use, answer quality and safety.

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “AWS open-sources UAEF, a tool to test AI agents across frameworks”

    Comments on “AWS open-sources UAEF, a tool to test AI agents across frameworks”
  99. No. 99Score 0.56

    BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

    Summary and score for “BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model”

    Tether AI Research has released BotPsy, a 0.8B-parameter open-weight model built on Qwen3.5-0.8B for function calling and agentic tasks on personal devices. The company reports it roughly doubles its base model's tool-use scores and outperforms all tested models of up to 1.2B parameters on six agentic benchmarks, though it trails…

    Score 0.56Contestedjudges differ by 0.43 on reader interest

    Elsewhere: HN —

    Score breakdown for “BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model”

    Comments on “BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model”
  100. No. 100Score 0.57

    Adventures in Weight Space Programming

    Summary and score for “Adventures in Weight Space Programming”

    Beaker, an autonomous agent improver, was given the task from MIT CSAIL researchers Takehi et al. of editing a language model's weights directly, without training, to fix its loops on MATH-500.

    Score 0.57

    Elsewhere: HN 7 pts, front page

    Score breakdown for “Adventures in Weight Space Programming”

    Comments on “Adventures in Weight Space Programming”
  101. No. 101Score 0.55

    Attention to Mamba: A Recipe for Cross-Architecture Distillation

    Summary and score for “Attention to Mamba: A Recipe for Cross-Architecture Distillation”

    Researchers Abhinav Moudgil, Ningyuan Huang, and colleagues propose a two-stage recipe for distilling a Transformer into a Mamba-style state space model without keeping any attention blocks. The method first linearizes attention via a kernel trick, then distills that into Mamba.

    Score 0.55

    Elsewhere: HN 69 pts, front page

    Score breakdown for “Attention to Mamba: A Recipe for Cross-Architecture Distillation”

    Comments on “Attention to Mamba: A Recipe for Cross-Architecture Distillation”
  102. No. 102Score 0.55

    Spared – self-hosted Workers and Durable Objects for Rust apps

    Summary and score for “Spared – self-hosted Workers and Durable Objects for Rust apps”

    Spared is an open-source, self-hosted runtime for Rust web apps, offered as a single binary that runs many apps, each with its own SQLite database, inside separate WebAssembly sandboxes. It aims to match Cloudflare Workers and Durable Objects, with stronger data isolation.

    Score 0.55

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “Spared – self-hosted Workers and Durable Objects for Rust apps”

    Comments on “Spared – self-hosted Workers and Durable Objects for Rust apps”
  103. No. 103Score 0.55

    Axiomitisation of Scienctific Notations for Programming Languages

    Summary and score for “Axiomitisation of Scienctific Notations for Programming Languages”

    The author argues that typeclasses (as in Haskell and Rust) and module systems (as in OCaml) are often confused because each serves a different purpose: typeclasses provide ad-hoc polymorphism for reusing operator names across types, while modules provide modular abstraction for structuring large programs.

    Score 0.55

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Axiomitisation of Scienctific Notations for Programming Languages”

    Comments on “Axiomitisation of Scienctific Notations for Programming Languages”
  104. No. 104Score 0.55

    What a Coding Agent Costs

    Summary and score for “What a Coding Agent Costs”

    A developer built smelt, an open-source coding agent, and wrote a benchmark measuring CPU and memory across seven agents using a scripted model for identical runs on the same hardware. Memory at idle ranged from 5.8 MiB (fx) to 430 MiB (opencode).

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “What a Coding Agent Costs”

    Comments on “What a Coding Agent Costs”
  105. No. 105Score 0.55

    [$] Additional hardening for the glibc dynamic linker

    Summary and score for “[$] Additional hardening for the glibc dynamic linker”

    The GNU C Library is considering new tunable settings that would let administrators forbid programs lacking address space layout randomization (ASLR) support. The proposal follows a Node.js Brotli decoder buffer overflow, where official Linux builds could not enable ASLR, making remote code execution easier.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “[$] Additional hardening for the glibc dynamic linker”

    Comments on “[$] Additional hardening for the glibc dynamic linker”
  106. No. 106Score 0.59

    Open research, stronger privacy: Making Firefox users harder to fingerprint

    Summary and score for “Open research, stronger privacy: Making Firefox users harder to fingerprint”

    Mozilla reports new Firefox protections against browser fingerprinting, the technique of identifying users across sites by combining details like fonts, screen size, and graphics hardware without cookies.

    Score 0.59

    Elsewhere: HN 171 pts, front page

    Score breakdown for “Open research, stronger privacy: Making Firefox users harder to fingerprint”

    Comments on “Open research, stronger privacy: Making Firefox users harder to fingerprint”

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 window1,243
Scored blind by two judges1,153
Off topic233
Read by the third judge237
Claim-checked25
Published106