Top stories

  1. Atlassian patches a pre-auth file read in eight products, six of them Data Center editions, and a public exploit chain turns it into Jira administrator Atlassian patched a pre-auth file read in eight products, six Data Center editions, and watchTowr reached Jira admin.
  2. Tensorlake npm package shipped a credential worm that wipes the home directory if the stolen token is revoked Tensorlake npm 0.5.144 installs a service that runs rm -rf on the home directory when the stolen GitHub token dies.
  3. curl pulls 8.23.0 forward to 2026-10-14 for 22 vulnerabilities, including its first HIGH severity CVE since 2023 curl moved release 8.23.0 to 2026-10-14 to fix 22 vulnerabilities, one rated HIGH, with no detail until that day.
  4. OpenAI withdraws three of its mathematics manuscripts over a sign error and revises 14 more OpenAI withdrew three of its 2026-10-06 mathematics manuscripts over a sign error and revised 14 others.
  5. Margaret Hamilton, who led the Apollo flight software and began using the term software engineering, died at 90 Margaret Hamilton, who led the Apollo onboard flight software and began using the term software engineering, died at 90.

AI

  1. OpenAI announces GPT-6 and model-composed interactive UI in ChatGPT OpenAI announced GPT-6 and ChatGPT answers that compose visuals and interactive elements, with details unverified here.
  2. Association for Human Mathematics urges mathematicians to stop working with OpenAI The Association for Human Mathematics asked mathematicians to stop collaborating with OpenAI over its math release.
  3. Cactus releases Whistle, a 16.9 MB on-device speech recognition model, with its benchmark method stated Cactus released Whistle, a 16.9 MB speech recognition model, against 145.3 MB for Whisper base, CPU only.

ML research

  1. Paper argues Lean verification of AI autoformalisation proves nothing about the prose proof, and that OpenAI's Navier-Stokes case shows it A paper argues a passing Lean check says nothing about the prose proof it was translated from, citing OpenAI's release.

Agentic coding

  1. Docker Agent defines agents in YAML and distributes them through any OCI registry Docker Agent, an Apache-2.0 docker CLI plugin, defines agents in YAML and distributes them via any OCI registry.
  2. Microsoft says Copilot will route tasks to an on-device model by end of month, and sandbox tool calls in OS containers Microsoft says Copilot will run MAI Code 1.1 Flash on RTX Spark PCs by month end and sandbox tools via OS containers.

Security

  1. MCP trust gaps let one internal agent pass injected instructions to another at five unrelated organisations Researchers passed injected instructions between internal agents at five organisations through MCP trust assumptions.
  2. ShinyHunters bypassed the published PeopleSoft WAF mitigation with URL encoding before extorting a divested Boeing unit ShinyHunters defeated Mandiant's PeopleSoft WAF rules with URL encoding and extorted a divested Boeing unit.
  3. FakeGit re-aimed 17,610 existing GitHub repositories at a SmartLoader download in 34 hours The FakeGit campaign re-aimed 17,610 existing GitHub repositories at a SmartLoader download in 34 hours.

Languages and runtimes

  1. Bun merges a built-in TypeScript type checker that matches tsc 7 errors across 1,102 of 1,103 project configs Bun merged a built-in TypeScript type checker that matched tsc 7 errors across 1,102 of 1,103 project configs.
  2. LLM-written Rust port of the TypeScript compiler passes all 181,711 ported Go tests and type checks in about half of tsc 7's time An LLM-written Rust port of tsc passes all 181,711 ported Go tests and type checks in about half of tsc 7's time.

Linux and kernel

  1. LWN report gives the kernel maintainer figures behind the LLM bug-report flood: 33 CVEs a day and a mean time to exploit of 7 days LWN reports kernel maintainers now fix 33 CVEs a day, and mean time from bug to exploit fell from 63 days to 7.

Engineering posts

  1. incident.io traces a Postgres queue ceiling to MultiXact lock contention caused by per-item savepoints incident.io found per-item savepoints put a MultiXact in every locked row, capping its Postgres queue throughput.
  2. Push ifs up and fors down, with the algebra that says when each rewrite is legal A post gives the preconditions that make pushing conditions up and loops down a legal rewrite rather than a style rule.
  3. GitHub reports one in three pull requests involves an AI agent and no detectable trend in the per-push secret leak rate GitHub reports one in three pull requests involves an AI agent and no detectable trend in the per-push secret leak rate.