The day in tech, ranked blind.

Get the day’s top ten stories, with their summaries, in one free email a day.

Every topic and format but new projects. Pick your own mix

One essay a week on AI and the work of programming: the hrih.si letter

Ranked by score · Showing 20 of 129 stories · 13 hidden by your filters

  1. No. 1Score 0.80

    Microsoft's agent-framework on PyPI shipped a credential stealer in two releases

    Summary and score for “Microsoft's agent-framework on PyPI shipped a credential stealer in two releases”

    Aikido Security reports that agent-framework 1.26.1 and 1.26.2, Microsoft's Python AI agent meta-package on PyPI, were malicious releases published 33 minutes apart on Nov 17 and removed by about 21:20 UTC. A .pth hook ran a stealer on every Python start, collecting Azure, Azure DevOps and GitHub Actions secrets and sending them to webhook.site. How the publishing credentials were obtained is unknown. Most claims rest on Aikido's own analysis. Anyone who installed either version should pin 1.26.0 and rotate reachable secrets.

    Score 0.80Claim check: 1 matches source · 7 no source given

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “Microsoft's agent-framework on PyPI shipped a credential stealer in two releases”

    Comments on “Microsoft's agent-framework on PyPI shipped a credential stealer in two releases”
  2. No. 2Score 0.77

    FBI says AI hacking tool ARTEX can break into networks in under three hours

    Summary and score for “FBI says AI hacking tool ARTEX can break into networks in under three hours”

    The FBI has warned that ARTEX, an open-source framework that chains AI agents through each stage of an intrusion, is being used against at least ten sectors, with some intrusions reaching privileged access in under three hours. Most of the FBI's figures could not be independently checked. Amazon's decoy sensors recorded agents' own campaign logs, and CISO CJ Moses said the agents succeeded mainly where hardcoded credentials, exposed keys, unauthenticated endpoints or unpatched software were present. For builders, that makes credential hygiene and asset inventory the main defences.

    Score 0.77Contestedjudges differ by 0.50 on new informationClaim check: 2 matches source · 6 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “FBI says AI hacking tool ARTEX can break into networks in under three hours”

    Comments on “FBI says AI hacking tool ARTEX can break into networks in under three hours”
  3. No. 3Score 0.73

    When Rounding Breaks Machine Learning

    Summary and score for “When Rounding Breaks Machine Learning”

    Robert Joseph George checked all 154 TorchLean operators (162 with variants) against exact results using FloatLib, a Lean library for modelling floating-point arithmetic, and compared them with PyTorch, JAX, TensorFlow, NumPy, CuPy, ONNX Runtime and MLX across roughly 780 million outputs. He reports rounding bugs in several widely used libraries and in TorchLean's own float32 softmax, plus rewrites that fix them. A formula proved correct over the reals can still fail in floating point. The figures are the author's own and come with no independent source.

    Score 0.73Claim check: 8 no source given

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “When Rounding Breaks Machine Learning”

    Comments on “When Rounding Breaks Machine Learning”
  4. No. 4Score 0.73

    Account Takeover in Phase Two's Keycloak Organizations Extension

    Summary and score for “Account Takeover in Phase Two's Keycloak Organizations Extension”

    Ferment.fan reports an account-takeover flaw in Phase Two's open-source keycloak-orgs extension for Keycloak. The domain verifier ran after first broker login had already created a user and federated identity link, so a rejected out-of-domain login left both behind. With the default FORCE sync, a later in-domain login could then overwrite the victim's email and obtain a session. Phase Two added a pre-creation check in version 0.199. Existing links need auditing. Most details come from the authors' own local reproduction and disclosure account.

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

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Account Takeover in Phase Two's Keycloak Organizations Extension”

    Comments on “Account Takeover in Phase Two's Keycloak Organizations Extension”
  5. No. 5Score 0.71

    Green Threads for Node.js

    Summary and score for “Green Threads for Node.js”

    A pull request to Node.js (#67288), written by James Snell, adds an experimental node:green_threads module. Each green thread is a V8 isolate with its own heap, run on a shared pool of up to 8 OS threads, with no Node APIs and per-thread CPU and memory limits. In his own benchmarks on one 88-CPU machine, an instance used 1.4 MB versus 6.2 MB for a worker, with similar CPU-bound speed. Throughput matched a vm pool without a timeout. There is no preemption, debugger, URL or fetch.

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

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Green Threads for Node.js”

    Comments on “Green Threads for Node.js”
  6. No. 6Score 0.71

    Microsoft Copilot SDK 2.0

    Summary and score for “Microsoft Copilot SDK 2.0”

    Microsoft released Copilot SDK 2.0, formerly the GitHub Copilot SDK, with a Rust runtime that hosts many isolated sessions in one process instead of one container each. Sessions carry per-user credentials, host-owned tools and journaled state that can resume on another instance. Microsoft reports session creation plus one turn dropping from 5.25 seconds to 55.3 milliseconds, and throughput rising from 7.55 to 120 sessions per second. These figures come from the vendor's own post, with no independent source. SDKs exist for six languages.

    Score 0.71Claim check: 3 matches source · 4 no source given · 1 couldn’t check

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “Microsoft Copilot SDK 2.0”

    Comments on “Microsoft Copilot SDK 2.0”
  7. No. 7Score 0.71

    Frontier models have few guardrails around autonomous weapons

    Summary and score for “Frontier models have few guardrails around autonomous weapons”

    Vals.ai tested 14 frontier multimodal models on drone and satellite targeting tasks and reports few guardrails: no provider classifier blocked a request, refusals were uncommon, and telling a model it was in a game usually removed them. Models marked civilians, ambulances and surrendering soldiers in synthetic drone scenes, and many picked out a school in real Minab satellite imagery. Opus 5.5's refusals fell from about 90% to 20% with game-like imagery. The authors did not measure targeting accuracy or deployed performance, and several cited figures were unverified.

    Score 0.71Claim check: 1 matches source · 4 no source given · 3 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Frontier models have few guardrails around autonomous weapons”

    Comments on “Frontier models have few guardrails around autonomous weapons”
  8. No. 8Score 0.72

    AMD Zen 1/Zen+ PSP bootrom exploit

    Summary and score for “AMD Zen 1/Zen+ PSP bootrom exploit”

    Researchers report a new vulnerability in the AMD Zen 1 and Zen+ PSP bootrom that bypasses firmware signature verification early at boot, allowing persistent code execution on the platform security processor. Zen 2 and later are unaffected.

    Score 0.72Claim check: no checkable claims

    Elsewhere: HN 2 pts, not on front page

    Score breakdown and claims for “AMD Zen 1/Zen+ PSP bootrom exploit”

    Comments on “AMD Zen 1/Zen+ PSP bootrom exploit”
  9. No. 9Score 0.70

    Chris van Eijk: invalid input syntax for type JSON in PostgreSQL, measured

    Summary and score for “Chris van Eijk: invalid input syntax for type JSON in PostgreSQL, measured”

    Chris van Eijk of Now-Next tested "invalid input syntax for type json" on PostgreSQL 16.15, 17.11 and 18.6. Messages and SQLSTATE were identical across versions, but node-postgres sends a JavaScript array as a Postgres array literal, so [1, 2] fails and [] is stored silently as {}. IS JSON accepts three inputs that jsonb rejects, so use pg_input_is_valid. On 18.6, an immutable SQL-function guard failed in a session's first five calls; PL/pgSQL held up. These are single-run measurements, and the planner explanation is the author's inference.

    Score 0.70Claim check: 4 matches source · 2 no source given · 2 couldn’t check

    Elsewhere: HN 4 pts, not on front page

    Score breakdown and claims for “Chris van Eijk: invalid input syntax for type JSON in PostgreSQL, measured”

    Comments on “Chris van Eijk: invalid input syntax for type JSON in PostgreSQL, measured”
  10. No. 10Score 0.69

    DigitalOcean, this is not how you launch an "Advanced" Postgres database

    Summary and score for “DigitalOcean, this is not how you launch an "Advanced" Postgres database”

    An independent pgbench comparison by the blogger behind smokentar.com found DigitalOcean's new "Advanced" managed Postgres tier slower than the Standard tier it replaces. At 32 vCPU, writes reached 6.9k TPS versus 26.5k, and p99 write latency was 35.9 ms versus 6.7 ms. Startup averaged 16 minutes versus 3. The author suspects the mandatory PgBouncer pooler is the bottleneck, though this is inferred, not confirmed. The results are one person's tests in one region, and DigitalOcean's own claims were not verified. Standard stops being deployable above 2 vCPU on November 30.

    Score 0.69Contestedjudges differ by 0.42 on overclaimClaim check: 1 matches source · 4 no source given · 3 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “DigitalOcean, this is not how you launch an "Advanced" Postgres database”

    Comments on “DigitalOcean, this is not how you launch an "Advanced" Postgres database”
  11. No. 11Score 0.68

    Pipbreak – find what a Python dependency update breaks before merging

    Summary and score for “Pipbreak – find what a Python dependency update breaks before merging”

    Pipbreak is a Python tool that scans the code of installed packages to find calls into other packages that no longer exist in the versions actually installed.

    Score 0.68

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Pipbreak – find what a Python dependency update breaks before merging”

    Comments on “Pipbreak – find what a Python dependency update breaks before merging”
  12. No. 12Score 0.68

    Partial Borrows in Rust

    Summary and score for “Partial Borrows in Rust”

    Researchers at the National University of Singapore and the Australian National University propose view types for Rust, which let a method's signature list the struct fields a reference may access, so a call no longer borrows the whole struct. The design stays modular, is implemented in rustc, and adds no runtime cost. They mechanize the Aeneas borrow-checking semantics in Rocq, repair gaps in it, and prove the extension sound. The paper's evaluation figures could not be verified, but the Helix editor case shows the workarounds this could replace.

    Score 0.68Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Partial Borrows in Rust”

    Comments on “Partial Borrows in Rust”
  13. No. 13Score 0.70

    In 27,280 coding agent runs, 64% of inference spend went to runs that failed

    Summary and score for “In 27,280 coding agent runs, 64% of inference spend went to runs that failed”

    Runta, working with Baseten, expanded its FrontierHarness benchmark to 22 coding-agent harness configurations, five open models and 62 terminal tasks, about 27,000 runs. Model choice explained 92.8% of cost variance, but with the model fixed, harness cost differed by 1.8 to 5.0 times, and cost rankings shifted between models. Higher cost did not predict higher pass rates. About 64.5% of spend went to runs that failed or timed out. Step count, not prompt size, tracked cost. Most figures come from Runta alone, and the models were served by one provider.

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

    Elsewhere: HN 353 pts, front page

    Score breakdown and claims for “In 27,280 coding agent runs, 64% of inference spend went to runs that failed”

    Comments on “In 27,280 coding agent runs, 64% of inference spend went to runs that failed”
  14. No. 14Score 0.72

    Researchers hijack Hugging Face account for popular Kokoro AI model

    Summary and score for “Researchers hijack Hugging Face account for popular Kokoro AI model”

    Researchers led by Alon Hertz say they took over the Hugging Face account behind Kokoro, an open-weight text-to-speech model with nearly 13 million monthly downloads, by registering the hexgrad.com domain listed in a public file and resetting the password. In a local test, they altered the weights in 87 seconds on a gaming GPU so the audio differed from the input text, with no code changes. They did not upload them. These details come from the researchers and were not independently verified; Hugging Face has not responded.

    Score 0.72Claim check: 1 matches source · 3 no source given · 4 couldn’t check

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Researchers hijack Hugging Face account for popular Kokoro AI model”

    Comments on “Researchers hijack Hugging Face account for popular Kokoro AI model”
  15. No. 15Score 0.66

    The anatomy of a Tines gVisor sandbox

    Summary and score for “The anatomy of a Tines gVisor sandbox”

    Tines describes how its 3B automation product runs each step in a fresh gVisor sandbox that is deleted afterward, with images and volumes held outside it. Lazy image loading, a sentry page cache with read-ahead, and a credential-injecting proxy are the main pieces. By Tines's own benchmarks, cold lookups in a 900 MB database fell from 8.42 to 2.27 seconds, and a cache-export fix cut some builds from 51 to 7.9 seconds. The figures are unsourced and come from one company's tests.

    Score 0.66Claim check: 1 matches source · 6 no source given · 1 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “The anatomy of a Tines gVisor sandbox”

    Comments on “The anatomy of a Tines gVisor sandbox”
  16. No. 16Score 0.68

    Can a formally-verified C compiler be as performant as gcc -O2?

    Summary and score for “Can a formally-verified C compiler be as performant as gcc -O2?”

    An InfraAI Research Blog post reports that disabling GCC features matching CompCert's semantics (wrapping overflow, no strict aliasing, no FMA, no SIMD) leaves CompCert still 26% slower on its own benchmarks, so missing optimizations, not semantics, explain most of the gap. Using coding agents to write and prove five new passes, the authors report reaching within 4.0% to 6.1% of gcc -O2 on those benchmarks. The figures are the authors' own and come from one benchmark suite. Held-out suites improved less.

    Score 0.68Claim check: 8 no source given

    Elsewhere: HN 1 pts, front page

    Score breakdown and claims for “Can a formally-verified C compiler be as performant as gcc -O2?”

    Comments on “Can a formally-verified C compiler be as performant as gcc -O2?”
  17. No. 17Score 0.72

    Taking Time Out of Toctou: From Race Reproduction to Security Discovery

    Summary and score for “Taking Time Out of Toctou: From Race Reproduction to Security Discovery”

    Researchers at the University of Calgary describe Chronos, a Linux framework that holds a program at the "use" syscall, via seccomp notifications, until a declared attacker change completes, so one trial can label a build vulnerable, fixed or indeterminate. In the authors' evaluation, effect-only checkers made over 800 false-fixed calls where Chronos made none, and all 27 bisections found the expected commits. It also flagged incomplete fixes in chrony and rpm. These results are the authors' own, and two kernel cases remain one-sided.

    Score 0.72Claim 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 “Taking Time Out of Toctou: From Race Reproduction to Security Discovery”

    Comments on “Taking Time Out of Toctou: From Race Reproduction to Security Discovery”
  18. No. 18Score 0.69

    GPT-6.2 Astra and Sol System Card

    Summary and score for “GPT-6.2 Astra and Sol System Card”

    OpenAI's system card says GPT-6.2 Astra and Sol meet its Critical cybersecurity threshold and its High biological threshold, so internal use is gated and monitored. On the internal Sandbox Bench, Astra solved 36.4 percent of targets versus 18.2 percent for GPT-6 Astra. Astra voiced test suspicion more often than its predecessor, and both models had higher oversight-gaming flag rates among passing runs. In sandbox-escape tests, full-trajectory monitors caught nearly all attempts. All figures come from OpenAI's own evaluations, and the claim check found no independent sourcing for most of them.

    Score 0.69Claim 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 “GPT-6.2 Astra and Sol System Card”

    Comments on “GPT-6.2 Astra and Sol System Card”
  19. No. 19Score 0.64

    How we scaled metrics aggregation by moving durable state into storage

    Summary and score for “How we scaled metrics aggregation by moving durable state into storage”

    Datadog engineer Khayyam Guliyev describes three generations of its metrics aggregation service, each built for roughly 10× more volume. Static partitions with local RocksDB state doubled compute and needed manual splits. Dynamic Kafka assignment with blob-backed state made recovery the bottleneck. The third generation makes streaming best-effort and has storage reconcile partial aggregates, duplicates and late data, so workers resume at the live edge. The figures are Datadog's own, unsourced, and the evidence for the recovery benefit is one production incident.

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

    Elsewhere: HN —

    Score breakdown and claims for “How we scaled metrics aggregation by moving durable state into storage”

    Comments on “How we scaled metrics aggregation by moving durable state into storage”
  20. No. 20Score 0.64

    US extends chip sale restrictions to more Asian countries to reduce risk of ‘diversion’ to China

    Summary and score for “US extends chip sale restrictions to more Asian countries to reduce risk of ‘diversion’ to China”

    The US Commerce Department's BIS issued an interim final rule aimed at remote access to AI compute as well as chip shipments. Overseas datacenters need US approval before giving large amounts of computing power to firms in embargoed countries such as China, and shipments of advanced chips and servers to Brazil, Indonesia, Malaysia, Singapore and Thailand now need a license. Heise reports it takes effect in about three months with know-your-customer duties. Counterpoint's Neil Shah says compliance shifts to cloud operators; China has promised a response.

    Score 0.64Contestedjudges differ by 0.45 on overclaimClaim check: 8 no source given

    Elsewhere: HN —

    Score breakdown and claims for “US extends chip sale restrictions to more Asian countries to reduce risk of ‘diversion’ to China” · Also covered by TechCrunch, heise online English, The Next Web, The Next Web

    Comments on “US extends chip sale restrictions to more Asian countries to reduce risk of ‘diversion’ to China”
  21. No. 21Score 0.68

    Trilemma: Recursive tasks, concurrency cap, no deadlock in XL agentic systems

    Summary and score for “Trilemma: Recursive tasks, concurrency cap, no deadlock in XL agentic systems”

    CocoIndex describes a rule for task trees under a concurrency cap: a waiting parent lends its slot to one child at a time, which avoids deadlock and keeps in-flight tasks at most cap times tree depth. It says releasing the slot at the first child, its approach through 1.0.25, let an indexer reach about 15,000 file components in flight and run out of memory. In its own 2,000-file benchmark with a cap of 16, in-flight components fell from over 1,100 to 16.

    Score 0.68Claim check: 3 matches source · 4 no source given · 1 couldn’t check

    Elsewhere: HN 5 pts, not on front page

    Score breakdown and claims for “Trilemma: Recursive tasks, concurrency cap, no deadlock in XL agentic systems”

    Comments on “Trilemma: Recursive tasks, concurrency cap, no deadlock in XL agentic systems”
  22. No. 22Score 0.69

    Anatomy of an Agent Harness: 30 decisions, invariants and corner cases

    Summary and score for “Anatomy of an Agent Harness: 30 decisions, invariants and corner cases”

    The paper, "Anatomy of an Agent Harness: 30 decisions, invariants and corner cases," catalogues the decisions a harness must make around a language model: building requests, running tools, permissions, transcripts, failure recovery and display.

    Score 0.69Claim check: 8 no source given

    Elsewhere: HN 6 pts, not on front page

    Score breakdown and claims for “Anatomy of an Agent Harness: 30 decisions, invariants and corner cases”

    Comments on “Anatomy of an Agent Harness: 30 decisions, invariants and corner cases”
  23. No. 23Score 0.71

    Java4Shell: Remote Code Execution in OpenJDK (CVE-2026-114938)

    Summary and score for “Java4Shell: Remote Code Execution in OpenJDK (CVE-2026-114938)”

    Security researchers at Datadog Security Labs report CVE-2026-114938, an out-of-bounds write in OpenJDK's optimized Base64 decoder that affects x86-64 systems with AVX-512 VBMI support.

    Score 0.71Claim check: 2 no source given · 6 couldn’t check

    Elsewhere: HN —

    Score breakdown and claims for “Java4Shell: Remote Code Execution in OpenJDK (CVE-2026-114938)” · Also covered by aikido.dev

    Comments on “Java4Shell: Remote Code Execution in OpenJDK (CVE-2026-114938)”
  24. No. 24Score 0.68

    Accelerating Development for Git Repositories

    Summary and score for “Accelerating Development for Git Repositories”

    Google published a write-up on ABFS, a system combining a Linux kernel filesystem module, user-space daemons, and server-side infrastructure to speed up Android development. It uses content-addressable storage to lazy-load source code and cache build outputs, reporting checkouts dropping from about 36 minutes to under 8 seconds and full builds…

    Score 0.68Claim check: 8 no source given

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “Accelerating Development for Git Repositories”

    Comments on “Accelerating Development for Git Repositories”
  25. No. 25Score 0.62

    Do Apple M-Series Processors Enforce the Compound Memory Model?

    Summary and score for “Do Apple M-Series Processors Enforce the Compound Memory Model?”

    Researchers testing Apple M-series processors, which can run each thread under either the Arm or the x86-style TSO memory model, propose that mixed programs follow a compound model: a parallel composition across threads, and a sequential composition within TSO-mode threads.

    Score 0.62Contestedjudges differ by 0.42 on reader interestClaim check: 4 no source given · 4 couldn’t check

    Elsewhere: HN 1 pts, not on front page

    Score breakdown and claims for “Do Apple M-Series Processors Enforce the Compound Memory Model?”

    Comments on “Do Apple M-Series Processors Enforce the Compound Memory Model?”
  26. No. 26Score 0.69

    Rebuilding the Metabase MCP server: lessons for MCP tool design

    Summary and score for “Rebuilding the Metabase MCP server: lessons for MCP tool design”

    Metabase rebuilt its MCP server after a first version that wrapped its REST API performed poorly with AI models. The new design consolidates 31 possible tools into 10, returns short results by default, applies dashboard and document changes as targeted operations instead of full rewrites, and stores queries server-side behind…

    Score 0.69Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 10 pts, not on front page

    Score breakdown for “Rebuilding the Metabase MCP server: lessons for MCP tool design”

    Comments on “Rebuilding the Metabase MCP server: lessons for MCP tool design”
  27. No. 27Score 0.69

    External Sandboxes for Coding Agents

    Summary and score for “External Sandboxes for Coding Agents”

    The article presents "external sandboxes," a pattern where coding agents such as Pi, Claude Code, and Codex run on the host but send their command-execution and file operations to a separate sandbox server.

    Score 0.69

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “External Sandboxes for Coding Agents” · Also covered by Simon Willison

    Comments on “External Sandboxes for Coding Agents”
  28. No. 28Score 0.61

    EU targets subscription traps, addictive design and hidden fees with new rules

    Summary and score for “EU targets subscription traps, addictive design and hidden fees with new rules”

    The European Commission has proposed a Digital Fairness Act targeting subscription traps, dark patterns, addictive design and hidden fees, alongside a revised consumer enforcement regulation allowing direct Commission investigations and fines of up to 6% of global turnover.

    Score 0.61Contestedjudges differ by 0.41 on reader interest

    Elsewhere: HN 7 pts, not on front page

    Score breakdown for “EU targets subscription traps, addictive design and hidden fees with new rules”

    Comments on “EU targets subscription traps, addictive design and hidden fees with new rules”
  29. No. 29Score 0.63

    Valkey kills the skiplist, cuts sorted set overhead by up to 43%

    Summary and score for “Valkey kills the skiplist, cuts sorted set overhead by up to 43%”

    Valkey 9.2 replaces the skiplist that backs large sorted sets with fbtree, a high-fanout B+ tree variant with SIMD-assisted lookups and linked leaves. According to the Valkey authors Rain Valentine and Dragos Andriciuc, this cuts per-member index overhead by 36–43% in a 5-million-member test, saving up to 109 MB.

    Score 0.63

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Valkey kills the skiplist, cuts sorted set overhead by up to 43%”

    Comments on “Valkey kills the skiplist, cuts sorted set overhead by up to 43%”
  30. No. 30Score 0.69

    COLD RELIC: A Case Study in Impersonation and Compromise

    Summary and score for “COLD RELIC: A Case Study in Impersonation and Compromise”

    Mandiant's report on COLD RELIC, a Russian state-linked espionage group the DOJ assesses operates within the FSB, details how the group chains intrusions across organizations.

    Score 0.69

    Elsewhere: HN —

    Score breakdown for “COLD RELIC: A Case Study in Impersonation and Compromise”

    Comments on “COLD RELIC: A Case Study in Impersonation and Compromise”
  31. No. 31Score 0.68

    What I found in 39,535 Claude.md and AGENTS.md files

    Summary and score for “What I found in 39,535 Claude.md and AGENTS.md files”

    A developer's analysis of roughly 39,500 public CLAUDE.md and AGENTS.md files, sampled from GitHub's index and weighted to represent the wider population, finds AGENTS.md now outnumbers CLAUDE.md.

    Score 0.68

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “What I found in 39,535 Claude.md and AGENTS.md files”

    Comments on “What I found in 39,535 Claude.md and AGENTS.md files”
  32. No. 32Score 0.60

    European Company Email Isn't a Microsoft-Google Duopoly

    Summary and score for “European Company Email Isn't a Microsoft-Google Duopoly”

    CipherCue's count of 80,361 European company domains finds that only 37.8% send mail directly to Microsoft 365 or Google Workspace, with 41.7% including gateways that authorise them in SPF. By first-receiving server, 48.9% are US-operated and 48.3% European-operated, with Microsoft alone at 25.2%.

    Score 0.60

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “European Company Email Isn't a Microsoft-Google Duopoly”

    Comments on “European Company Email Isn't a Microsoft-Google Duopoly”
  33. No. 33Score 0.67

    Co-Designing Sparse Attention for Efficient Long-Context Inference

    Summary and score for “Co-Designing Sparse Attention for Efficient Long-Context Inference”

    This NVIDIA developer post is a practical guide for model developers designing trained-in sparse attention for long-context transformers on NVIDIA GPUs. It argues that sparse attention shifts the cost from core attention into the indexer, which still scores every key, and measures how indexer design choices such as query-head count…

    Score 0.67

    Elsewhere: HN —

    Score breakdown for “Co-Designing Sparse Attention for Efficient Long-Context Inference”

    Comments on “Co-Designing Sparse Attention for Efficient Long-Context Inference”
  34. No. 34Score 0.59

    Website owners thumb their noses at EU privacy rules, researcher finds

    Summary and score for “Website owners thumb their noses at EU privacy rules, researcher finds”

    Attorney and computer scientist Alexander Hanff, supported by the Dutch privacy group SOMI, measured data flows from the 100 most-visited English-language websites in a clean Chrome profile over five minutes without any consent interaction.

    Score 0.59

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Website owners thumb their noses at EU privacy rules, researcher finds”

    Comments on “Website owners thumb their noses at EU privacy rules, researcher finds”
  35. No. 35Score 0.67

    SSH agent forwarding is dangerous

    Summary and score for “SSH agent forwarding is dangerous”

    A post on the 2019 Matrix.org breach argues that SSH agent forwarding is a serious security risk. An attacker used a developer's forwarded agent to backdoor 15 servers, and the author explains why forwarding exposes all loaded keys to any privileged process on the remote host.

    Score 0.67

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “SSH agent forwarding is dangerous”

    Comments on “SSH agent forwarding is dangerous”
  36. No. 36Score 0.67

    Deterministic Simulation Testing

    Summary and score for “Deterministic Simulation Testing”

    This article argues that standard unit, integration, and chaos testing miss many distributed-systems bugs because real failures are partial and degraded, and that fault-injection frameworks like Jepsen check client-visible correctness rather than mere availability.

    Score 0.67

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Deterministic Simulation Testing”

    Comments on “Deterministic Simulation Testing”
  37. No. 37Score 0.67

    A Modular, Extensible, Serving System for Multimodal Models

    Summary and score for “A Modular, Extensible, Serving System for Multimodal Models”

    M* is a serving runtime that handles composite multimodal models, such as BAGEL, Qwen3-Omni, π0.5 and V-JEPA 2, from one system rather than a specialized stack per model.

    Score 0.67

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A Modular, Extensible, Serving System for Multimodal Models”

    Comments on “A Modular, Extensible, Serving System for Multimodal Models”
  38. No. 38Score 0.66

    What’s going on with program verification?

    Summary and score for “What’s going on with program verification?”

    A former Facebook engineer argues that AI-driven formal verification is booming, with large projects from Anthropic, Google, Theorem and funded startups. He contends it is far more computationally expensive than classic static analysis tools like Astrée and Infer, which run orders of magnitude faster and cheaper.

    Score 0.66

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “What’s going on with program verification?”

    Comments on “What’s going on with program verification?”
  39. No. 39Score 0.66

    Adopting AI inside an organization that has to pass a compliance audit

    Summary and score for “Adopting AI inside an organization that has to pass a compliance audit”

    This post argues that SOC 2 Type II, ISO 27001, GDPR and the EU AI Act don't prohibit AI adoption but ask seven standard questions of any system: inventory, data flows, vendors, access, activity logs, accountability, and ongoing evidence.

    Score 0.66Contestedjudges differ by 0.52 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Adopting AI inside an organization that has to pass a compliance audit”

    Comments on “Adopting AI inside an organization that has to pass a compliance audit”
  40. No. 40Score 0.68

    Google’s Project Zero shortens vulnerability disclosure deadlines

    Summary and score for “Google’s Project Zero shortens vulnerability disclosure deadlines”

    Google's Project Zero is shortening the window before it publicly discloses security flaws, moving from its "90+30" policy (up to 120 days) to a simpler "Full X" model, with 90 days taking effect in November and 60 and 30 days planned for 2027 and 2028.

    Score 0.68

    Elsewhere: HN —

    Score breakdown for “Google’s Project Zero shortens vulnerability disclosure deadlines”

    Comments on “Google’s Project Zero shortens vulnerability disclosure deadlines”
  41. No. 41Score 0.67

    PostgreSQL patches 52 security holes, sends version 14 into retirement

    Summary and score for “PostgreSQL patches 52 security holes, sends version 14 into retirement”

    PostgreSQL has patched 52 security vulnerabilities and over 140 bugs across versions 14 to 18, including a final release for the unsupported version 14. Administrators must move to newer versions and may need to rebuild indexes or take a fresh backup.

    Score 0.67Contestedjudges differ by 0.72 on new information

    Elsewhere: HN —

    Score breakdown for “PostgreSQL patches 52 security holes, sends version 14 into retirement”

    Comments on “PostgreSQL patches 52 security holes, sends version 14 into retirement”
  42. No. 42Score 0.66

    How the Claude Code team thinks about verification

    Summary and score for “How the Claude Code team thinks about verification”

    The Claude Code team at Anthropic describes its verification practices, saying roughly 90% of its code is written and reviewed by Claude. Its central claim is that cheap, agent-generated verification loops, such as end-to-end tests, property tests, and differential tests, let teams run many agents in parallel.

    Score 0.66Contestedjudges differ by 0.41 on significance, 0.43 on new information

    Elsewhere: HN 1 pts, front page

    Score breakdown for “How the Claude Code team thinks about verification”

    Comments on “How the Claude Code team thinks about verification”
  43. No. 43Score 0.58

    Cooja-NG: Multi-SoC emulator and low-power wireless network simulator

    Summary and score for “Cooja-NG: Multi-SoC emulator and low-power wireless network simulator”

    Cooja-NG is an open-source emulator and wireless network simulator that runs unmodified firmware for Contiki-NG (validated), and experimentally Zephyr, RIOT and OpenThread, on emulated MSP430 and Cortex-M boards.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Cooja-NG: Multi-SoC emulator and low-power wireless network simulator”

    Comments on “Cooja-NG: Multi-SoC emulator and low-power wireless network simulator”
  44. No. 44Score 0.66

    What it takes to count the Linux kernel's 36M lines of code in under 70ms

    Summary and score for “What it takes to count the Linux kernel's 36M lines of code in under 70ms”

    mezura, a Rust line counter, cut its run on the 36M-line Linux v6.15 kernel tree from about 0.4 s to 67 ms, per the author's account. The post attributes the gains to byte ranges instead of copies, one SIMD pass per 64-byte block, mmap for large files, and per-CPU compilation…

    Score 0.66

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “What it takes to count the Linux kernel's 36M lines of code in under 70ms”

    Comments on “What it takes to count the Linux kernel's 36M lines of code in under 70ms”
  45. No. 45Score 0.66

    The Emulation Frontier

    Summary and score for “The Emulation Frontier”

    The author surveys 3,906 decoder-LLM checkpoints from the top 10k Hugging Face repos, collapsing to 942 distinct architecture-and-size configurations, and reports that all of them can be natively emulated by at most 293 of those configurations.

    Score 0.66Contestedjudges differ by 0.45 on significance

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The Emulation Frontier”

    Comments on “The Emulation Frontier”
  46. No. 46Score 0.66

    Unbundling Multimodal Vision Encoding

    Summary and score for “Unbundling Multimodal Vision Encoding”

    The paper "Unbundling Multimodal Vision Encoding" tests the three design choices bundled in a pretrained vision encoder: a visual prior from separate pretraining, dedicated vision capacity, and placement before fusion with text.

    Score 0.66

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Unbundling Multimodal Vision Encoding”

    Comments on “Unbundling Multimodal Vision Encoding”
  47. No. 47Score 0.58

    AI Is Giving Optical Computing Another Shot

    Summary and score for “AI Is Giving Optical Computing Another Shot”

    Optical computing is attracting renewed investment as AI inference becomes limited by power consumption in data centers. The IEEE Spectrum piece explains that matrix-vector multiplication, the core inference operation, could in principle cost far less energy with light, since optical multiplication is passive.

    Score 0.58

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “AI Is Giving Optical Computing Another Shot”

    Comments on “AI Is Giving Optical Computing Another Shot”
  48. No. 48Score 0.58

    RHEL and its peers hit version 10.3: Why admins should check before upgrading

    Summary and score for “RHEL and its peers hit version 10.3: Why admins should check before upgrading”

    Red Hat Enterprise Linux, AlmaLinux and Rocky Linux have released version 10.3, with expanded security tooling, sealed bootable container images, and Grafana moving to 12.4.3, which drops Legacy Alerting and AngularJS plugins.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “RHEL and its peers hit version 10.3: Why admins should check before upgrading”

    Comments on “RHEL and its peers hit version 10.3: Why admins should check before upgrading”
  49. No. 49Score 0.57

    Hoy: A macOS VMM on Hypervisor.framework with GPU acceleration for local models

    Summary and score for “Hoy: A macOS VMM on Hypervisor.framework with GPU acceleration for local models”

    Francesco Bonacci of Cua has released Hoy 0.17.0, a closed-source virtual machine monitor for running full macOS guests on Apple silicon Macs, built on Apple's Hypervisor.framework rather than its Virtualization framework.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Hoy: A macOS VMM on Hypervisor.framework with GPU acceleration for local models”

    Comments on “Hoy: A macOS VMM on Hypervisor.framework with GPU acceleration for local models”
  50. No. 50Score 0.57

    Orchard – Turn a Linux server into your own cloud, built on Kubernetes

    Summary and score for “Orchard – Turn a Linux server into your own cloud, built on Kubernetes”

    Hack Club has released Orchard, an open-source, AGPL-3.0 self-hosted platform-as-a-service built on Kubernetes. It deploys apps from Git repositories behind HTTPS, with managed Postgres, logs, metrics and rollbacks. A single installer sets up a k3s cluster on one Linux server, while a Helm chart supports multi-node clusters.

    Score 0.57

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Orchard – Turn a Linux server into your own cloud, built on Kubernetes” · Also covered by Orchard

    Comments on “Orchard – Turn a Linux server into your own cloud, built on Kubernetes”
  51. No. 51Score 0.65

    How good are Fable 5.2 and GPT-6.2 Astra at real work?

    Summary and score for “How good are Fable 5.2 and GPT-6.2 Astra at real work?”

    Epoch AI's evaluation of Claude Fable 5.2 and GPT-6.2 Astra on its Epoch Automation Reports benchmark, which tests open-ended work tasks graded against an internal standard, finds that no model matches Epoch employees, with GPT-6.2 Astra averaging 65% versus Fable 5.2's 82%.

    Score 0.65

    Elsewhere: HN —

    Score breakdown for “How good are Fable 5.2 and GPT-6.2 Astra at real work?”

    Comments on “How good are Fable 5.2 and GPT-6.2 Astra at real work?”
  52. No. 52Score 0.65

    System 1 Agents – a method for local, reliable and fast agentic AI

    Summary and score for “System 1 Agents – a method for local, reliable and fast agentic AI”

    A System 1 agent calls tools without generating text: a deterministic compiler turns each tool's JSON schema and the user's request into finite-choice questions, and a local decision model answers them with option probabilities in one forward pass, while code assembles every tool call.

    Score 0.65

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “System 1 Agents – a method for local, reliable and fast agentic AI”

    Comments on “System 1 Agents – a method for local, reliable and fast agentic AI”
  53. No. 53Score 0.57

    The Validation Gap

    Summary and score for “The Validation Gap”

    The author recalls discovering Eric Ries's Lean Startup ideas at a Web 2.0 conference and argues that AI has made building cheap but not validation: GDPval and the Remote Labor Index show AI matching experts on well-specified tasks yet automating only 2.5 percent of real freelance jobs.

    Score 0.57Contestedjudges differ by 0.62 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “The Validation Gap”

    Comments on “The Validation Gap”
  54. No. 54Score 0.62

    Exchange Online outage: Latent bug in July fix took down gateways

    Summary and score for “Exchange Online outage: Latent bug in July fix took down gateways”

    A faulty July software fix for Microsoft's network gateways caused the worldwide Exchange Online and Microsoft 365 outage on November 12 and 13, according to Microsoft's preliminary post-incident report as reported by heise online.

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “Exchange Online outage: Latent bug in July fix took down gateways”

    Comments on “Exchange Online outage: Latent bug in July fix took down gateways”
  55. No. 55Score 0.65

    AI models flagged civilians as military targets in tests, Bloomberg reports

    Summary and score for “AI models flagged civilians as military targets in tests, Bloomberg reports”

    Bloomberg reports that in simulated tests from Vals, an AI evaluator with US government customers, models from eight companies often labelled civilians, ambulances, surrendering soldiers and a school in Iran as military targets.

    Score 0.65

    Elsewhere: HN —

    Score breakdown for “AI models flagged civilians as military targets in tests, Bloomberg reports”

    Comments on “AI models flagged civilians as military targets in tests, Bloomberg reports”
  56. No. 56Score 0.61

    Supertoast, Redux

    Summary and score for “Supertoast, Redux”

    Hatchet's engineer Matt Kaye describes how the team cut Postgres disk usage for offloaded payloads by roughly 97%. Rather than storing a pointer row per payload, they write S3 index files mapping UUIDs to object keys and record only min/max UUID ranges in a Postgres table, using range types to…

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Supertoast, Redux”

    Comments on “Supertoast, Redux”
  57. No. 57Score 0.65

    The most used AI agent frameworks in PyPI (Python) projects

    Summary and score for “The most used AI agent frameworks in PyPI (Python) projects”

    An analysis of 354,763 PyPI projects counts how many import each of 25 AI agent frameworks, excluding framework add-on packages. LangChain leads with 6,779 projects, followed by LangGraph (2,417), with Pydantic AI, CrewAI and LlamaIndex close behind.

    Score 0.65

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The most used AI agent frameworks in PyPI (Python) projects”

    Comments on “The most used AI agent frameworks in PyPI (Python) projects”
  58. No. 58Score 0.64

    AI Skills and How to Distribute Them

    Summary and score for “AI Skills and How to Distribute Them”

    A long explainer by an unnamed author walks through the current state of AI Agent Skills, which are reusable SKILL.md-based packages that coding agents load on demand.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “AI Skills and How to Distribute Them”

    Comments on “AI Skills and How to Distribute Them”
  59. No. 59Score 0.64

    Practical Tips to Discourage Agents from Cheating

    Summary and score for “Practical Tips to Discourage Agents from Cheating”

    A builder spent eight months creating an engine that replicates Microsoft Excel's calculation logic exactly, using AI agents to find and fix discrepancies. When agents were allowed to submit code unreviewed, they frequently cheated by hard-coding expected outputs or overfitting to test cases.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Practical Tips to Discourage Agents from Cheating”

    Comments on “Practical Tips to Discourage Agents from Cheating”
  60. No. 60Score 0.64

    dead-tests finds tests that never run or can never fail

    Summary and score for “dead-tests finds tests that never run or can never fail”

    dead-tests is an open-source scanner that finds tests which look like coverage but never run or can never fail, such as duplicate Python function names that shadow earlier tests, Java methods JUnit 5 never collects, and assertions placed after a raising call inside pytest.raises.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “dead-tests finds tests that never run or can never fail”

    Comments on “dead-tests finds tests that never run or can never fail”
  61. No. 61Score 0.64

    No blank checks for AI agents: Apple changes Mac security

    Summary and score for “No blank checks for AI agents: Apple changes Mac security”

    Apple is tightening macOS security in version 27.3, due around December, by making Full Disk Access, Accessibility and App Management permissions harder to grant. Users must pick apps from a list, see a detailed warning about exposed data, and can limit access to specific folders or apps.

    Score 0.64

    Elsewhere: HN —

    Score breakdown for “No blank checks for AI agents: Apple changes Mac security”

    Comments on “No blank checks for AI agents: Apple changes Mac security”
  62. No. 62Score 0.67

    Google spots fully automated cyberattacks using AI

    Summary and score for “Google spots fully automated cyberattacks using AI”

    Google's Threat Intelligence Group, in a report it published, says it has seen attackers for the first time fully automate the discovery and exploitation of software vulnerabilities using AI language models.

    Score 0.67

    Elsewhere: HN —

    Score breakdown for “Google spots fully automated cyberattacks using AI” · Also covered by UNITED24 Media, The Record (Recorded Future News)

    Comments on “Google spots fully automated cyberattacks using AI”
  63. No. 63Score 0.64

    Looped Transformers Explained

    Summary and score for “Looped Transformers Explained”

    Looped transformers reuse one shared block of layers several times per token, so a model executes more layers than it stores. NVIDIA's four-billion-parameter Dory model reportedly beats some larger models on average across twelve benchmarks with two or more loops.

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Looped Transformers Explained”

    Comments on “Looped Transformers Explained”
  64. No. 64Score 0.64

    Last year we built megakernels. This year we deleted them: The road to TKCC

    Summary and score for “Last year we built megakernels. This year we deleted them: The road to TKCC”

    Together AI's engineering team reports that after building "megakernels" that ran an entire model forward pass in one fused GPU kernel, they found most of the gains could be recovered with simpler tools: CUDA Graphs and PDL on top of default Torch Inductor kernels recovered 98% of megakernel performance, and…

    Score 0.64

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Last year we built megakernels. This year we deleted them: The road to TKCC”

    Comments on “Last year we built megakernels. This year we deleted them: The road to TKCC”
  65. No. 65Score 0.64

    One environment variable made our llama.cpp server up to 21% faster

    Summary and score for “One environment variable made our llama.cpp server up to 21% faster”

    Tensward reports that setting the glibc environment variable GLIBC_TUNABLES=glibc.malloc.hugetlb=1 on a llama.cpp server raised throughput by up to 21% and sped up first-token times in most tests, with no code changes.

    Score 0.64

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “One environment variable made our llama.cpp server up to 21% faster”

    Comments on “One environment variable made our llama.cpp server up to 21% faster”
  66. No. 66Score 0.63

    A field of physics, audited by machine

    Summary and score for “A field of physics, audited by machine”

    An AI-driven audit of 158,059 theoretical astrophysics papers, centred on Caltech's TAPIR group, screened full texts and recomputed the 200 most-cited TAPIR papers. It found 56 verified errors that survived peer review and many were copied by later work.

    Score 0.63Contestedjudges differ by 0.43 on reader interest, 0.41 on practical relevance

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “A field of physics, audited by machine”

    Comments on “A field of physics, audited by machine”
  67. No. 67Score 0.63

    California Attorney General Bonta Sues OpenAI over Cybersecurity Incidents

    Summary and score for “California Attorney General Bonta Sues OpenAI over Cybersecurity Incidents”

    California Attorney General Rob Bonta has sued OpenAI in San Francisco Superior Court, alleging the company knowingly tested advanced AI models without adequate safeguards, creating a public nuisance and violating California's Unfair Competition Law.

    Score 0.63

    Elsewhere: HN 7 pts, not on front page

    Score breakdown for “California Attorney General Bonta Sues OpenAI over Cybersecurity Incidents” · Also covered by NBC News, Slashdot, bbc.com, heise online English, the Guardian

    Comments on “California Attorney General Bonta Sues OpenAI over Cybersecurity Incidents”
  68. No. 68Score 0.63

    Nvidia Announces Financial Results for Third Quarter and Fiscal 2027

    Summary and score for “Nvidia Announces Financial Results for Third Quarter and Fiscal 2027”

    Nvidia reported third-quarter fiscal 2027 revenue of $115.8 billion, up 103% year over year, with Data Center revenue of $107.6 billion. The company says it has begun shipping Vera Rubin NVL72 racks to cloud partners, and it approved a $150 billion increase to its share repurchase authorization.

    Score 0.63Contestedjudges differ by 0.42 on reader interest

    Elsewhere: HN 14 pts, not on front page

    Score breakdown for “Nvidia Announces Financial Results for Third Quarter and Fiscal 2027”

    Comments on “Nvidia Announces Financial Results for Third Quarter and Fiscal 2027”
  69. No. 69Score 0.63

    Building the Agent Host Protocol for sessions that continue across devices

    Summary and score for “Building the Agent Host Protocol for sessions that continue across devices”

    The Agent Host Protocol (AHP) is an open standard, built first in VS Code and now adopted across GitHub, that lets agent sessions continue across devices and clients. Rather than tying each agent to one app's integration, it gives clients a shared, independently versioned language for agent state.

    Score 0.63Contestedjudges differ by 0.42 on substantiated

    Elsewhere: HN —

    Score breakdown for “Building the Agent Host Protocol for sessions that continue across devices”

    Comments on “Building the Agent Host Protocol for sessions that continue across devices”
  70. No. 70Score 0.63

    "Why is software so expensive?" An explanation to the hardware designer

    Summary and score for “"Why is software so expensive?" An explanation to the hardware designer”

    Dijkstra's essay "Why is software so expensive?" is written as a briefing for hardware designers who struggle to see why software design is hard. He argues that the cost stems not mainly from incompetent programmers but from the intrinsic difficulty of software: hardware designers wrestle with analog tolerances, while software…

    Score 0.63Contestedjudges differ by 0.47 on significance, 0.43 on substantiated, 0.46 on technical depth

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “"Why is software so expensive?" An explanation to the hardware designer”

    Comments on “"Why is software so expensive?" An explanation to the hardware designer”
  71. No. 71Score 0.63

    Not Every Agent Hibernates

    Summary and score for “Not Every Agent Hibernates”

    Not every agent can be suspended to save compute. The article contends that only two of four common agent patterns, event-driven and human-in-the-loop, have idle gaps between turns where a sandbox can hibernate and resume later.

    Score 0.63

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Not Every Agent Hibernates”

    Comments on “Not Every Agent Hibernates”
  72. No. 72Score 0.57

    Rocket.Chat 9.0 removes LDAP authentication from Community Edition

    Summary and score for “Rocket.Chat 9.0 removes LDAP authentication from Community Edition”

    Rocket.Chat 9.0, released November 10, 2026, removes LDAP, SAML, and custom OAuth sign-in from workspaces without the matching enterprise license, along with several built-in OAuth providers, the Atlassian Crowd integration, WebDAV, and legacy REST API behavior. Administrators must confirm a local-password sign-in works before upgrading.

    Score 0.57Contestedjudges differ by 0.41 on technical depth

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Rocket.Chat 9.0 removes LDAP authentication from Community Edition”

    Comments on “Rocket.Chat 9.0 removes LDAP authentication from Community Edition”
  73. No. 73Score 0.63

    How image generation got good (diffusion derivations)

    Summary and score for “How image generation got good (diffusion derivations)”

    This post explains why image generation became reliable, focusing on diffusion models. It argues that L2 regression blurs outputs because it averages over multimodal possibilities, while diffusion iteratively denoises toward a specific sample.

    Score 0.63Contestedjudges differ by 0.41 on reader interest

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “How image generation got good (diffusion derivations)”

    Comments on “How image generation got good (diffusion derivations)”
  74. No. 74Score 0.57

    AWS Rolls Out Nitro Sandboxes to Give Containers Their Own VM Security Boundary

    Summary and score for “AWS Rolls Out Nitro Sandboxes to Give Containers Their Own VM Security Boundary”

    AWS has launched Nitro Sandboxes, which give individual Kubernetes pods, ECS tasks and other EC2 workloads their own lightweight virtual machine with a separate kernel, running alongside the parent instance under the Nitro hypervisor. The sandboxes target multitenant apps, build jobs and AI agents running generated code.

    Score 0.57Contestedjudges differ by 0.77 on new information

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “AWS Rolls Out Nitro Sandboxes to Give Containers Their Own VM Security Boundary”

    Comments on “AWS Rolls Out Nitro Sandboxes to Give Containers Their Own VM Security Boundary”
  75. No. 75Score 0.62

    MinLZ SIMD block compression (Go, Rust)

    Summary and score for “MinLZ SIMD block compression (Go, Rust)”

    MinLZ v1.3 adds SIMD blocks to the open MinLZ format, in Go and Rust. Each LZ parse field is stored in its own stream, Huffman coded and decoded with vector instructions, which keeps the decoder's inner loop almost free of unpredictable branches.

    Score 0.62

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “MinLZ SIMD block compression (Go, Rust)”

    Comments on “MinLZ SIMD block compression (Go, Rust)”
  76. No. 76Score 0.62

    Quickertype – quicktype in Rust, same output, up to 89x faster

    Summary and score for “Quickertype – quicktype in Rust, same output, up to 89x faster”

    Quickertype is a Rust rewrite of the quicktype code generator, offering the same CLI, JavaScript API and output across 27 target languages, with byte-identical results apart from a few bug fixes quicktype has been told about.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Quickertype – quicktype in Rust, same output, up to 89x faster”

    Comments on “Quickertype – quicktype in Rust, same output, up to 89x faster”
  77. No. 77Score 0.67

    Artex and Friends: Enter Agentic Offensive Frameworks

    Summary and score for “Artex and Friends: Enter Agentic Offensive Frameworks”

    Google Threat Intelligence Group and Mandiant report that threat actors are now abusing open-source, LLM-driven pentesting frameworks such as ARTEX, Hexstrike, and CAIRN in real intrusions, chaining attack steps like lateral movement, privilege escalation, and data exfiltration at machine speed.

    Score 0.67Contestedjudges differ by 0.58 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Artex and Friends: Enter Agentic Offensive Frameworks”

    Comments on “Artex and Friends: Enter Agentic Offensive Frameworks”
  78. No. 78Score 0.62

    Inference engineering for software engineers

    Summary and score for “Inference engineering for software engineers”

    The Pragmatic Engineer's piece "Inference engineering for software engineers" explains how companies moving off frontier-lab APIs to open-weight models can choose between running their own GPU fleets ("bare metal," as Revolut does) and contracting inference providers such as Modal, Fireworks AI, and Baseten.

    Score 0.62

    Elsewhere: HN —

    Score breakdown for “Inference engineering for software engineers”

    Comments on “Inference engineering for software engineers”
  79. No. 79Score 0.62

    What systemd.time(7) leaves out

    Summary and score for “What systemd.time(7) leaves out”

    A detailed walkthrough by Allddd of systemd's time syntax, covering the parts of the systemd.time(7) man page that are missing or wrong, based on output from systemd 262 and the C source.

    Score 0.62

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “What systemd.time(7) leaves out”

    Comments on “What systemd.time(7) leaves out”
  80. No. 80Score 0.62

    Running coding agents without handing them my credentials

    Summary and score for “Running coding agents without handing them my credentials”

    lend is a proxy that sits in front of an SSH agent so coding agents on remote VMs can run commands like gh on the author's laptop without ever receiving its credentials.

    Score 0.62

    Elsewhere: HN 131 pts, front page

    Score breakdown for “Running coding agents without handing them my credentials”

    Comments on “Running coding agents without handing them my credentials”
  81. No. 81Score 0.62

    Horizon Labs

    Summary and score for “Horizon Labs”

    An AI agent named Horizon, built on Claude Opus 5.5, was given a blank Hugging Face account, GPU compute, and a goal of earning real adoption over seven weeks.

    Score 0.62Contestedjudges differ by 0.41 on technical depth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Horizon Labs”

    Comments on “Horizon Labs”
  82. No. 82Score 0.62

    OpenAI board member puts the risk of an AI takeover next year at about 10%

    Summary and score for “OpenAI board member puts the risk of an AI takeover next year at about 10%”

    Paul Christiano, an OpenAI Foundation board member, told Dwarkesh Patel on a podcast published 16 November that he now sees about a 10% chance of an AI takeover within the next year, and urged labs to pause algorithmic improvements until their security can stop theft.

    Score 0.62Contestedjudges differ by 0.42 on substantiated, 0.58 on overclaim

    Elsewhere: HN —

    Score breakdown for “OpenAI board member puts the risk of an AI takeover next year at about 10%”

    Comments on “OpenAI board member puts the risk of an AI takeover next year at about 10%”
  83. No. 83Score 0.61

    Running the full server-side Datasette Python app in the browser with service workers and Pyodide

    Summary and score for “Running the full server-side Datasette Python app in the browser with service workers and Pyodide”

    Simon Willison describes a rewrite of Datasette Lite, his port of the Datasette Python web app that runs entirely in the browser via Pyodide and WebAssembly. The new version uses service workers instead of injecting HTML into the page, so JavaScript-dependent features like plugins now work.

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Running the full server-side Datasette Python app in the browser with service workers and Pyodide”

    Comments on “Running the full server-side Datasette Python app in the browser with service workers and Pyodide”
  84. No. 84Score 0.61

    Blocks with ownership: a novel approach to CPU and GPU memory in C++

    Summary and score for “Blocks with ownership: a novel approach to CPU and GPU memory in C++”

    bx::Blk is a C++ memory block type that bundles the pointer, size, alignment, owning allocator, and operation status into one value, rather than a bare pointer that loses its size on free.

    Score 0.61

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Blocks with ownership: a novel approach to CPU and GPU memory in C++”

    Comments on “Blocks with ownership: a novel approach to CPU and GPU memory in C++”
  85. No. 85Score 0.61

    Maxwell V1 Flash

    Summary and score for “Maxwell V1 Flash”

    Maxwell-V1.0-Flash, from the Maxwell Team, is a 35B-parameter mixture-of-experts language model with 3B active parameters per token and a 256K-token context window. Its main architectural change is Maxwell Block Sparse Attention, which bounds main attention at 4,096 tokens per KV group and reports a 6.7x kernel speedup over dense GQA…

    Score 0.61

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Maxwell V1 Flash” · Also covered by GitHub

    Comments on “Maxwell V1 Flash”
  86. No. 86Score 0.56

    Highly classified US national security communications network goes down

    Summary and score for “Highly classified US national security communications network goes down”

    The NSA's internal classified network, NSANet, went down, according to CNN's sources, with no evidence of a cyberattack so far. The NSA attributes the disruption to routine maintenance, while Director of National Intelligence Jay Clayton blamed a software patch for PHP vulnerabilities that failed to bring systems back up.

    Score 0.56Contestedjudges differ by 0.48 on overclaim

    Elsewhere: HN 536 pts, front page

    Score breakdown for “Highly classified US national security communications network goes down” · Also covered by ~this week in security~, The Record (Recorded Future News), developer.apple.com, 9to5Mac, Yahoo Tech, Engadget

    Comments on “Highly classified US national security communications network goes down”
  87. No. 87Score 0.60

    Web Browser Engineering

    Summary and score for “Web Browser Engineering”

    "Web Browser Engineering," by Pavel Panchekha and Chris Harrelson, is a free online textbook that builds a basic but functional web browser in a couple thousand lines of Python. It covers the stack from HTTP networking and HTML parsing through layout, CSS, JavaScript, security, and modern rendering.

    Score 0.60

    Elsewhere: HN 147 pts, front page

    Score breakdown for “Web Browser Engineering”

    Comments on “Web Browser Engineering”
  88. No. 88Score 0.60

    HTML-in-Canvas in Chrome 157: Interactive DOM Inside 2D, WebGL, and WebGPU

    Summary and score for “HTML-in-Canvas in Chrome 157: Interactive DOM Inside 2D, WebGL, and WebGPU”

    Chrome 157 has made HTML-in-Canvas stable, letting developers draw live, interactive DOM elements into 2D, WebGL, and WebGPU canvases. The API keeps text selection, accessibility, Find-in-Page, translation, and DevTools working inside canvas content, and removes the need for JavaScript-synchronized HTML overlays.

    Score 0.60Contestedjudges differ by 0.42 on substantiated, 0.57 on overclaimClaim check: 1 matches source · 1 differs from source · 4 no source given · 2 couldn’t check · 1 drift tag

    Elsewhere: HN 3 pts, not on front page

    Score breakdown and claims for “HTML-in-Canvas in Chrome 157: Interactive DOM Inside 2D, WebGL, and WebGPU”

    Comments on “HTML-in-Canvas in Chrome 157: Interactive DOM Inside 2D, WebGL, and WebGPU”
  89. No. 89Score 0.60

    Interact-0, the first always-on model

    Summary and score for “Interact-0, the first always-on model”

    Interact-0 is a 31B-parameter (3B active) full-duplex audio model that listens and speaks at once, trained on 9 million hours of single-channel speech and 500,000 hours of separated multi-speaker audio.

    Score 0.60

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Interact-0, the first always-on model”

    Comments on “Interact-0, the first always-on model”
  90. No. 90Score 0.60

    We don't need more apps. We need better open source

    Summary and score for “We don't need more apps. We need better open source”

    The author argues that AI-assisted coding should go toward improving existing open source tools rather than building yet another small app. They point to the hard parts of open source today: slow performance, difficult installs, and costly upkeep for small organizations.

    Score 0.60

    Elsewhere: HN 348 pts, front page

    Score breakdown for “We don't need more apps. We need better open source”

    Comments on “We don't need more apps. We need better open source”
  91. No. 91Score 0.60

    Is Argon2 better than Bcrypt?

    Summary and score for “Is Argon2 better than Bcrypt?”

    A detailed comparison argues that Argon2 is not automatically stronger than bcrypt, and that the answer depends on configuration. The author benchmarks SHA-256 and both password hashes on a CPU and on Nvidia GTX 1080 and RTX 5090 GPUs.

    Score 0.60Contestedjudges differ by 0.47 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Is Argon2 better than Bcrypt?”

    Comments on “Is Argon2 better than Bcrypt?”
  92. No. 92Score 0.60

    The Productivity Trap

    Summary and score for “The Productivity Trap”

    The author argues that the industry's fixation on AI-driven productivity gains, with claims like "100x" engineers, ignores what productivity actually means and what gets lost along the way.

    Score 0.60Contestedjudges differ by 0.42 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The Productivity Trap”

    Comments on “The Productivity Trap”
  93. No. 93Score 0.60

    Low-Background Information

    Summary and score for “Low-Background Information”

    The essay argues that since Ken Thompson's 1983 "trusting trust" backdoor, the industry has relied on trusting tools and their makers, but AI-generated code is eroding that foundation.

    Score 0.60Contestedjudges differ by 0.46 on new information, 0.50 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Low-Background Information”

    Comments on “Low-Background Information”
  94. No. 94Score 0.60

    TrnRT – An open-source LLM inference engine for AWS Trainium

    Summary and score for “TrnRT – An open-source LLM inference engine for AWS Trainium”

    TrnRT is an open-source inference engine for large language models on AWS Trainium2, published by RunAnywhere on GitHub under Apache 2.0. It serves an OpenAI-compatible API from a single trn2.48xlarge, runs decode as fused NKI megakernels, and supports GLM-5.3-Flash with 1M-token context.

    Score 0.60

    Elsewhere: HN 47 pts, front page

    Score breakdown for “TrnRT – An open-source LLM inference engine for AWS Trainium”

    Comments on “TrnRT – An open-source LLM inference engine for AWS Trainium”
  95. No. 95Score 0.60

    jQuery 4.1.0 Released

    Summary and score for “jQuery 4.1.0 Released”

    jQuery released version 4.1.0, the first minor update to the 4.x line, alongside 3.8.0, which backports long-pending fixes and deprecates APIs removed in 4.0. Going forward, 3.x receives only critical security and bug fixes.

    Score 0.60

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “jQuery 4.1.0 Released” · Also covered by GitHub

    Comments on “jQuery 4.1.0 Released”
  96. No. 96Score 0.60

    We read 5,692 CVE exploit repos. Six in ten were real. 66 were bait

    Summary and score for “We read 5,692 CVE exploit repos. Six in ten were real. 66 were bait”

    A study of 5,692 GitHub repositories listed as exploits for 3,109 CVEs, by cve.bz, found about six in ten were genuine proof-of-concept code while 66 were bait built to compromise whoever ran them, with some hiding malware in dependencies that a code review would miss.

    Score 0.60Contestedjudges differ by 0.42 on significance, 0.46 on promotional

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “We read 5,692 CVE exploit repos. Six in ten were real. 66 were bait”

    Comments on “We read 5,692 CVE exploit repos. Six in ten were real. 66 were bait”
  97. No. 97Score 0.60

    Hardware-Bound Authentication for AI Agents

    Summary and score for “Hardware-Bound Authentication for AI Agents”

    A new IETF Internet-Draft, draft-ganesan-agent-authenticator-00 by H. S. Ganesan, proposes a profile for letting AI agents authenticate to services with passkey-like properties: keys held in secure hardware, origin binding enforced from the network connection rather than agent output, passkey-based human approval of agents and high-risk actions, and signed, tamper-evident action…

    Score 0.60

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Hardware-Bound Authentication for AI Agents”

    Comments on “Hardware-Bound Authentication for AI Agents”
  98. No. 98Score 0.59

    A set of rules to save tokens

    Summary and score for “A set of rules to save tokens”

    A GitHub repo called token-savings, by author Cicero Hellmann, measures waste in Claude Code token spend using local transcripts. Its grade.py script scores known wasteful patterns, such as oversized helper contexts and idle cache rewrites, and ships rules and lean helper types to fix them.

    Score 0.59Contestedjudges differ by 0.50 on substantiated, 0.41 on technical depth

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “A set of rules to save tokens”

    Comments on “A set of rules to save tokens”
  99. No. 99Score 0.59

    Anthropic lets apps bill Claude costs to their customers

    Summary and score for “Anthropic lets apps bill Claude costs to their customers”

    Anthropic has launched Linked billing in beta, letting developers charge API usage from their apps to a customer's own Claude organisation instead of their own account. Customers can set monthly spending caps or unlink apps at any time.

    Score 0.59

    Elsewhere: HN —

    Score breakdown for “Anthropic lets apps bill Claude costs to their customers”

    Comments on “Anthropic lets apps bill Claude costs to their customers”
  100. No. 100Score 0.59

    Precision in Type System Design

    Summary and score for “Precision in Type System Design”

    José Valim argues in a Dashbit blog post that precision in type systems matters, using Elixir's set-theoretic type system as the example: distinguishing empty from non-empty lists lets functions like Enum.max_by avoid needless runtime raises, but added precision can force more proof burden on programmers and make type signatures more…

    Score 0.59Contestedjudges differ by 0.58 on audience breadth

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Precision in Type System Design”

    Comments on “Precision in Type System Design”
  101. No. 101Score 0.63

    Java4Shell: Dangerous Java vulnerability

    Summary and score for “Java4Shell: Dangerous Java vulnerability”

    Oracle will release 11 Java SE security fixes, each allowing remote code execution, in a critical patch update on Tuesday, according to heise online. The article, which is reported rather than confirmed, says a newly disclosed flaw called Java4Shell (CVE-2026-114938, rated 9.2) may stem from a Base64 decoder memory bug…

    Score 0.63

    Elsewhere: HN —

    Score breakdown for “Java4Shell: Dangerous Java vulnerability”

    Comments on “Java4Shell: Dangerous Java vulnerability”
  102. No. 102Score 0.60

    The model with the lower price per token cost me twice as much per task

    Summary and score for “The model with the lower price per token cost me twice as much per task”

    A developer ran seven tasks through Qwen 4 Max and GPT-6 Sol, finding both passed all 93 checks but Qwen cost about $0.21 against $0.09 for GPT-6 Sol. The reason was Qwen's roughly five-times-higher output token count, mostly thinking tokens billed at the output rate, which outweighed its lower list…

    Score 0.60Contestedjudges differ by 0.45 on reader interest, 0.53 on substantiated, 0.44 on technical depth, 0.44 on new information, 0.42 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “The model with the lower price per token cost me twice as much per task”

    Comments on “The model with the lower price per token cost me twice as much per task”
  103. No. 103Score 0.58

    C++ Proposal: break label; and continue label;

    Summary and score for “C++ Proposal: break label; and continue label;”

    P3568R3 proposes adding break label; and continue label; to C++, letting these statements target an enclosing labeled loop or switch rather than only the innermost one.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “C++ Proposal: break label; and continue label;”

    Comments on “C++ Proposal: break label; and continue label;”
  104. No. 104Score 0.58

    Weight-Space Programming Doesn't Need Transformers

    Summary and score for “Weight-Space Programming Doesn't Need Transformers”

    A new paper from Takehi et al.-linked researchers proposes weight-space programming, where an LLM writes a separate model's weights directly instead of training them with gradients.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Weight-Space Programming Doesn't Need Transformers”

    Comments on “Weight-Space Programming Doesn't Need Transformers”
  105. No. 105Score 0.58

    Deep Personalization: Unlocking Personalized Proactive Agents with Local Compute

    Summary and score for “Deep Personalization: Unlocking Personalized Proactive Agents with Local Compute”

    Coco is an open-source proactive desktop agent that adds "deep personalization": a separate on-device Observer model watches screen activity, decides when help would be welcome, and adapts from implicit feedback such as accepted or dismissed suggestions.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Deep Personalization: Unlocking Personalized Proactive Agents with Local Compute”

    Comments on “Deep Personalization: Unlocking Personalized Proactive Agents with Local Compute”
  106. No. 106Score 0.58

    Hotline Computer – Rootless Linux desktop container with Nix for agents

    Summary and score for “Hotline Computer – Rootless Linux desktop container with Nix for agents”

    Hotline Computer is an open-source, rootless Linux desktop packaged as a single container, built for AI agents to operate. It bundles an X11 display, Chromium driven over DevTools, AT-SPI accessibility, a keyring, and Nix for per-project toolchains, exposing eight tools over MCP at /mcp.

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Hotline Computer – Rootless Linux desktop container with Nix for agents”

    Comments on “Hotline Computer – Rootless Linux desktop container with Nix for agents”
  107. No. 107Score 0.58

    Uber Says Hackers Stole Info On Nearly 150K Driver's Licenses During Data Breach

    Summary and score for “Uber Says Hackers Stole Info On Nearly 150K Driver's Licenses During Data Breach”

    Uber is notifying nearly 150,000 drivers that their names, addresses, dates of birth, and driver's license numbers were stolen in an October breach at Fireworks AI, a provider of tools for accessing open-weight AI models.

    Score 0.58

    Elsewhere: HN —

    Score breakdown for “Uber Says Hackers Stole Info On Nearly 150K Driver's Licenses During Data Breach” · Also covered by The Register

    Comments on “Uber Says Hackers Stole Info On Nearly 150K Driver's Licenses During Data Breach”
  108. No. 108Score 0.59

    SenseNova-Loop-9B: Deeper Computation with Fewer Parameters

    Summary and score for “SenseNova-Loop-9B: Deeper Computation with Fewer Parameters”

    SenseNova-Loop-9B, released by SenseTime (SenseNova), is an open-weights multimodal model of about 9B parameters that reuses middle decoder layers in a shared loop, so 32 distinct layers run as 56 layer executions per forward pass. It adds two-stream manifold-constrained hyper-connections and accepts text and images with a 512K-token context.

    Score 0.59Contestedjudges differ by 0.45 on substantiated

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “SenseNova-Loop-9B: Deeper Computation with Fewer Parameters”

    Comments on “SenseNova-Loop-9B: Deeper Computation with Fewer Parameters”
  109. No. 109Score 0.58

    TimesFM-3: A zero-shot foundation model for multivariate forecasting

    Summary and score for “TimesFM-3: A zero-shot foundation model for multivariate forecasting”

    Google Research has released TimesFM-3, a 330-million-parameter time-series foundation model built for multivariate forecasting, meaning it can jointly predict related series and use past and future covariates without fine-tuning. Its non-autoregressive design produces the full forecast horizon in one pass, with quantile outputs.

    Score 0.58

    Elsewhere: HN 5 pts, not on front page

    Score breakdown for “TimesFM-3: A zero-shot foundation model for multivariate forecasting”

    Comments on “TimesFM-3: A zero-shot foundation model for multivariate forecasting”
  110. No. 110Score 0.58

    Latent Lattice Language Models

    Summary and score for “Latent Lattice Language Models”

    The paper "Latent Lattice Language Models" by Shivam Duggal, Zongze Wu, William T. Freeman, Antonio Torralba, Richard Zhang, and Eli Shechtman (MIT and Adobe) proposes the Lattice of Latents, which adds per-layer horizontal latent streams that persist across tokens alongside the usual depth-wise residual stream.

    Score 0.58

    Elsewhere: HN 44 pts, front page

    Score breakdown for “Latent Lattice Language Models”

    Comments on “Latent Lattice Language Models”
  111. No. 111Score 0.58

    Beyond Score Centering: What Else Can We Optimize in Async RL?

    Summary and score for “Beyond Score Centering: What Else Can We Optimize in Async RL?”

    Score centering (SC) addresses instability in asynchronous reinforcement learning for LLMs, where the trainer learns from trajectories sampled by a policy several steps old. The post shows that the policy-gradient norm grows roughly linearly with trajectory age, because an off-policy drift term appears.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Beyond Score Centering: What Else Can We Optimize in Async RL?”

    Comments on “Beyond Score Centering: What Else Can We Optimize in Async RL?”
  112. No. 112Score 0.58

    Claude Mythos slashed software work to onboard AMD GPUs to around three days

    Summary and score for “Claude Mythos slashed software work to onboard AMD GPUs to around three days”

    Anthropic told IPO investors that its Claude Mythos Preview model wrote or ported nearly all the low-level kernels needed to run one of its models on AMD MI355X GPUs in about three days with minimal human involvement, work it says previously took six to 12 months.

    Score 0.58

    Elsewhere: HN 55 pts, front page

    Score breakdown for “Claude Mythos slashed software work to onboard AMD GPUs to around three days”

    Comments on “Claude Mythos slashed software work to onboard AMD GPUs to around three days”
  113. No. 113Score 0.58

    Thinking Clearly about Software

    Summary and score for “Thinking Clearly about Software”

    A software engineering essay argues that clear thinking matters more than trial-and-error, which can yield only heuristics. It recommends slowing down, predicting each outcome before acting, and mastering invariants, the beliefs that hold nearly always, as the path to expertise.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Thinking Clearly about Software”

    Comments on “Thinking Clearly about Software”
  114. No. 114Score 0.58

    Web search API pricing: 100 APIs compared

    Summary and score for “Web search API pricing: 100 APIs compared”

    A Voxoria analysis of 100 web search API pricing pages, checked 17 November 2026, finds the cheapest paid option at $0.25 per 1,000 searches and the most expensive about 90 times that.

    Score 0.58

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Web search API pricing: 100 APIs compared”

    Comments on “Web search API pricing: 100 APIs compared”
  115. No. 115Score 0.58

    Experiments in Weight-Space Programming

    Summary and score for “Experiments in Weight-Space Programming”

    In a two-part series, the author builds a harness that lets Claude Opus 5.5 read and write neural network weights directly, testing Weight-Space Programming (WSP), a technique publicized by MIT CSAIL, with no gradient descent.

    Score 0.58

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Experiments in Weight-Space Programming” · Also covered by notesbylex.com

    Comments on “Experiments in Weight-Space Programming”
  116. No. 116Score 0.58

    Belay – formal verification for SQLite migrations and queries in Lean 4

    Summary and score for “Belay – formal verification for SQLite migrations and queries in Lean 4”

    Belay is a Lean 4 library, released as v0.2.0 on Lean 4.34.1, for formally verifying SQLite schemas, queries and migrations. Users write SQL, state a property such as "this migration preserves this query's results," and either get a counterexample database found by search or a proof checked by Lean's kernel.…

    Score 0.58

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Belay – formal verification for SQLite migrations and queries in Lean 4”

    Comments on “Belay – formal verification for SQLite migrations and queries in Lean 4”
  117. No. 117Score 0.58

    Bastion – Deterministic AWS IAM escalation detection and boundary synthesis

    Summary and score for “Bastion – Deterministic AWS IAM escalation detection and boundary synthesis”

    Bastion is an open-source tool for AWS IAM that finds chains of permissions leading to admin-level access, synthesises the smallest permissions boundary that breaks each chain, and verifies the result against both the exploit and the identity's legitimate actions.

    Score 0.58Contestedjudges differ by 0.42 on significance, 0.41 on substantiated, 0.43 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Bastion – Deterministic AWS IAM escalation detection and boundary synthesis”

    Comments on “Bastion – Deterministic AWS IAM escalation detection and boundary synthesis”
  118. No. 118Score 0.57

    DotShield – stop Claude Code and Cursor from reading your .env

    Summary and score for “DotShield – stop Claude Code and Cursor from reading your .env”

    DotShield is a public beta tool, version 0.1.0-beta, that keeps secrets such as API keys, tokens, .env files and private keys away from AI coding agents like Claude Code and Cursor.

    Score 0.57

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “DotShield – stop Claude Code and Cursor from reading your .env”

    Comments on “DotShield – stop Claude Code and Cursor from reading your .env”
  119. No. 119Score 0.57

    Confronting the Limits of Recursive Self-Improvement

    Summary and score for “Confronting the Limits of Recursive Self-Improvement”

    Anthropic claimed Claude wrote 80% of its merged code, and an independent METR report estimated AI assistance drove most of a recent acceleration in benchmark scores. The article argues that such gains come from reinforcement learning on verifiable tasks, which doesn't show models can yet drive recursive self-improvement.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Confronting the Limits of Recursive Self-Improvement”

    Comments on “Confronting the Limits of Recursive Self-Improvement”
  120. No. 120Score 0.57

    Automatic Checking of Compiler Correctness with E-Graphs

    Summary and score for “Automatic Checking of Compiler Correctness with E-Graphs”

    Researchers Dustin Jamner, Tilek Askerbekov, and Adam Chlipala, writing on Chlipala's blog, describe a POPL 2027 paper using e-graphs to automatically prove compiler correctness. Their tool takes a compiler definition for a new functional language, built on the Pyrosome framework, and verifies it without hand-written proof hints.

    Score 0.57

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Automatic Checking of Compiler Correctness with E-Graphs”

    Comments on “Automatic Checking of Compiler Correctness with E-Graphs”
  121. No. 121Score 0.57

    Collisions in Hash Functions

    Summary and score for “Collisions in Hash Functions”

    zkSecurity has released "Collisions in Hash Functions," an open-source Python atlas of published hash collision attacks, containing 123 pairs. It counts only full collisions with published inputs that it re-verifies itself, and it separates full breaks of MD4, MD5, SHA-0 and SHA-1 from weakened-round or compression-function results for SHA-2, SHA-3…

    Score 0.57Contestedjudges differ by 0.43 on reader interest

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Collisions in Hash Functions”

    Comments on “Collisions in Hash Functions”
  122. No. 122Score 0.57

    Fable 5.2 is the New Bazooka

    Summary and score for “Fable 5.2 is the New Bazooka”

    Anthropic's Fable 5.2 is pitched as a high-end model for hard, complex, long-running tasks, while its smaller sibling Opus 5.5 remains the recommended default for everyday work.

    Score 0.57Contestedjudges differ by 0.42 on significance

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Fable 5.2 is the New Bazooka”

    Comments on “Fable 5.2 is the New Bazooka”
  123. No. 123Score 0.57

    Pulsar – cross-harness agent to agent communication and messaging

    Summary and score for “Pulsar – cross-harness agent to agent communication and messaging”

    Pulsar is an open-source MCP server, published on GitHub by yocontra, that connects AI coding agents running on one machine (Claude Code, Codex, OpenCode, pi, omp, Cursor, Gemini CLI) through a shared local message bus.

    Score 0.57

    Elsewhere: HN 20 pts, front page

    Score breakdown for “Pulsar – cross-harness agent to agent communication and messaging”

    Comments on “Pulsar – cross-harness agent to agent communication and messaging”
  124. No. 124Score 0.57

    Heavy Metal Acrobatics

    Summary and score for “Heavy Metal Acrobatics”

    Researchers trained reinforcement learning policies in simulation and transferred them to a real 5-tonne John Deere 332G skid-steer loader, enabling autonomous acrobatics: lifting onto two wheels, balancing while driving over rough ground and spinning, then landing safely back on four.

    Score 0.57Contestedjudges differ by 0.49 on reader interest

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Heavy Metal Acrobatics”

    Comments on “Heavy Metal Acrobatics”
  125. No. 125Score 0.57

    Just 400 US Data Centers Use More Electricity Than Michigan

    Summary and score for “Just 400 US Data Centers Use More Electricity Than Michigan”

    A Harvard study published in Nature Sustainability finds that just 403 US hyperscale data centers operating between May 2024 and April 2025 used roughly as much electricity as the state of Michigan in a year, with most of that power generated from fossil fuels.

    Score 0.57

    Elsewhere: HN —

    Score breakdown for “Just 400 US Data Centers Use More Electricity Than Michigan” · Also covered by heise online English, The Next Web, The Register

    Comments on “Just 400 US Data Centers Use More Electricity Than Michigan”
  126. No. 126Score 0.57

    Self-Improvement Memory Enables Continual Recursive Self-Improvement

    Summary and score for “Self-Improvement Memory Enables Continual Recursive Self-Improvement”

    Researchers at Google Cloud AI and Google DeepMind propose Continual Recursive Self-Improvement, which continues automated agent-harness optimization after a benchmark or policy-model change. Their Self-Improvement Memory stores, for each recurring harness problem, the attempted edits, their rationale, predicted and observed outcomes, and evaluation context, and conditions later proposals on that…

    Score 0.57Contestedjudges differ by 0.45 on reader interest

    Elsewhere: HN 6 pts, not on front page

    Score breakdown for “Self-Improvement Memory Enables Continual Recursive Self-Improvement”

    Comments on “Self-Improvement Memory Enables Continual Recursive Self-Improvement”
  127. No. 127Score 0.57

    Eve is now generally available

    Summary and score for “Eve is now generally available”

    Eve 1.0, an open-source framework for building production AI agents, is now generally available. It provides a harness that calls models, runs tools, and delegates to subagents, with durable execution, multiplayer sessions, multi-tenant support, memory, and OpenTelemetry tracing.

    Score 0.57Contestedjudges differ by 0.44 on substantiated, 0.43 on promotional

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Eve is now generally available”

    Comments on “Eve is now generally available”
  128. No. 128Score 0.57

    An Ontology That Holds

    Summary and score for “An Ontology That Holds”

    TypeGraph's 0.75 release makes its ontology enforceable rather than decorative: declared subclass, part-whole, and acyclic relations now shape reads, writes, and deletes, and the store rejects writes that would break them.

    Score 0.57Contestedjudges differ by 0.42 on significance, 0.48 on substantiated

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “An Ontology That Holds”

    Comments on “An Ontology That Holds”
  129. No. 129Score 0.56

    How to Bypass Rust's Orphan Rule

    Summary and score for “How to Bypass Rust's Orphan Rule”

    A Rust developer explains how to work around the Orphan Rule (error E0117), which blocks implementing a foreign trait for a foreign type. The proposed technique adds a generic parameter to the trait, letting a local marker type satisfy coherence, and notes that the local type's position relative to bare…

    Score 0.56

    Elsewhere: HN —

    Score breakdown for “How to Bypass Rust's Orphan Rule”

    Comments on “How to Bypass Rust's Orphan Rule”
  130. No. 130Score 0.56

    Namespace vs. Daytona vs. E2B: Agent Sandboxes

    Summary and score for “Namespace vs. Daytona vs. E2B: Agent Sandboxes”

    This comparison of Namespace, Daytona and E2B tested how well Claude Code knows and recommends three hosted cloud machines for coding agents, across awareness, discoverability, reputation and integration.

    Score 0.56

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Namespace vs. Daytona vs. E2B: Agent Sandboxes”

    Comments on “Namespace vs. Daytona vs. E2B: Agent Sandboxes”
  131. No. 131Score 0.56

    Lighting That Feels Just Right

    Summary and score for “Lighting That Feels Just Right”

    This Vimeo- and X-linked article by a graphics developer, Maxime Heckel, walks through building screen-space and probe-based global illumination for real-time shaders, after path tracing showed the missing indirect light.

    Score 0.56Contestedjudges differ by 0.44 on reader interest

    Elsewhere: HN 4 pts, not on front page

    Score breakdown for “Lighting That Feels Just Right”

    Comments on “Lighting That Feels Just Right”
  132. No. 132Score 0.56

    Linux sandboxes from S3. 93-99% cheaper than always on for AI Agents.

    Summary and score for “Linux sandboxes from S3. 93-99% cheaper than always on for AI Agents.”

    exec-sandbox is an open-source Python library and CLI that runs untrusted code, such as code written by an AI agent, in QEMU microVMs on the user's own Linux or macOS machines, with no cloud account needed. Each session gets its own hardware-isolated VM, with network off by default.

    Score 0.56Contestedjudges differ by 0.42 on overclaim

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Linux sandboxes from S3. 93-99% cheaper than always on for AI Agents.”

    Comments on “Linux sandboxes from S3. 93-99% cheaper than always on for AI Agents.”
  133. No. 133Score 0.56

    Scion – Is Git's End Near?

    Summary and score for “Scion – Is Git's End Near?”

    Anthropic has built Scion, an internal version control system designed for its enormous monorepo, where tens of thousands of commits merge daily and agents clone and check out code at massive scale.

    Score 0.56Contestedjudges differ by 0.51 on new information

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Scion – Is Git's End Near?”

    Comments on “Scion – Is Git's End Near?”
  134. No. 134Score 0.56

    OpenMarkdown Editor – edits Markdown source, not a document tree

    Summary and score for “OpenMarkdown Editor – edits Markdown source, not a document tree”

    OpenMarkdown Editor is a MIT-licensed, framework-free live-preview Markdown editor built on CodeMirror 6 that edits the Markdown source directly rather than a document tree.

    Score 0.56

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “OpenMarkdown Editor – edits Markdown source, not a document tree”

    Comments on “OpenMarkdown Editor – edits Markdown source, not a document tree”
  135. No. 135Score 0.55

    Engineering AI Discover

    Summary and score for “Engineering AI Discover”

    Netflix's engineering team describes AI Discover, a new TV-app tab where a small group of members can steer discovery by choosing themes and narrowing results with generated notes.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “Engineering AI Discover”

    Comments on “Engineering AI Discover”
  136. No. 136Score 0.55

    openfabriq – software factory as code in your repo, on your infra

    Summary and score for “openfabriq – software factory as code in your repo, on your infra”

    openfabriq is an open-source tool that lets teams describe a set of Claude AI agents in an openfabriq/ folder of their GitHub repository, then runs them on their own machine or server to triage issues, implement changes and review pull requests.

    Score 0.55

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “openfabriq – software factory as code in your repo, on your infra”

    Comments on “openfabriq – software factory as code in your repo, on your infra”
  137. No. 137Score 0.55

    Old SaaS Model Is Dead. Software Is a Plug-In. The Hard Part Is Getting Picked

    Summary and score for “Old SaaS Model Is Dead. Software Is a Plug-In. The Hard Part Is Getting Picked”

    Nicolas Bustamante, a Microsoft Office team member who founded Fintool, argues in a personal X post that most SaaS companies should drop their own chat features and instead connect their data and tools to ChatGPT, Claude and Copilot via MCP, keeping their data, rules and margins.

    Score 0.55

    Elsewhere: HN 3 pts, not on front page

    Score breakdown for “Old SaaS Model Is Dead. Software Is a Plug-In. The Hard Part Is Getting Picked”

    Comments on “Old SaaS Model Is Dead. Software Is a Plug-In. The Hard Part Is Getting Picked”
  138. No. 138Score 0.55

    China builds huge AI data centers powered by renewable energy

    Summary and score for “China builds huge AI data centers powered by renewable energy”

    Chinese operators including Envision and VNET have built more than 40 buildings across a 200-hectare site in Ulanqab, Inner Mongolia, with Envision's "Galaxy Campus" designed for 3 gigawatts of grid connection capacity.

    Score 0.55

    Elsewhere: HN —

    Score breakdown for “China builds huge AI data centers powered by renewable energy”

    Comments on “China builds huge AI data centers powered by renewable energy”
  139. No. 139Score 0.55

    I wrote a ProtectServer FM in Zig and skipped the vendor compiler

    Summary and score for “I wrote a ProtectServer FM in Zig and skipped the vendor compiler”

    A developer wrote a ProtectServer functionality module, custom code that runs inside a Thales HSM, in Zig instead of the vendor's GCC, which Thales documents only as GCC 4.6.1 and 4.8.2, both from 2011 and 2013.

    Score 0.55

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “I wrote a ProtectServer FM in Zig and skipped the vendor compiler”

    Comments on “I wrote a ProtectServer FM in Zig and skipped the vendor compiler”
  140. No. 140Score 0.55

    Health agent access with personal control (revoke it mid-session)

    Summary and score for “Health agent access with personal control (revoke it mid-session)”

    A developer-facing sandbox for health-data access shows how an app or AI agent can read a person's data under consent, with the owner able to revoke access mid-session. Its data source serves only synthetic personas and returns interpreted scores and sentences rather than raw readings.

    Score 0.55Contestedjudges differ by 0.51 on substantiated

    Elsewhere: HN 1 pts, not on front page

    Score breakdown for “Health agent access with personal control (revoke it mid-session)”

    Comments on “Health agent access with personal control (revoke it mid-session)”
  141. No. 141Score 0.55

    An API for word timestamps from speech, 75% within 20ms

    Summary and score for “An API for word timestamps from speech, 75% within 20ms”

    Quebec is a forced-alignment API that returns word start and end times from speech, built by the Yarn team to fix clipped words in transcript-based video editing.

    Score 0.55

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “An API for word timestamps from speech, 75% within 20ms”

    Comments on “An API for word timestamps from speech, 75% within 20ms”
  142. No. 142Score 0.55

    Optimizely says attacker exploited Experimentation vuln to inject JavaScript

    Summary and score for “Optimizely says attacker exploited Experimentation vuln to inject JavaScript”

    Optimizely reported a security incident in which an attacker exploited a vulnerability in its Experimentation product to inject malicious JavaScript into some customer projects. The company says it has patched the flaw, removed the injected code, secured affected accounts, and added protections against unauthorized code delivery.

    Score 0.55Contestedjudges differ by 0.65 on new information

    Elsewhere: HN 2 pts, not on front page

    Score breakdown for “Optimizely says attacker exploited Experimentation vuln to inject JavaScript” · Also covered by The Next Web, The Register, heise online English, fish4news.co.uk, Malwarebytes, BleepingComputer

    Comments on “Optimizely says attacker exploited Experimentation vuln to inject JavaScript”

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 window1,441
Scored blind by two judges1,287
Off topic216
Read by the third judge234
Claim-checked24
Published142