Sunday 18 October 2026

Top 3

1. OpenCode breach hits 1.4% of Console users

   OpenCode says a security breach tied to a large-scale distillation attack hit 1.4% of its Console users, and stolen API keys led to surprise usage bills for some. It urges affected users to revoke their keys and change passwords, and says it is working with Google, OpenAI and Anthropic.

   huggingnews.com

2. DeepSeek launches Search and Fetch APIs

   Two new endpoints give agents live web access: Search returns structured results from DeepSeek's own search engine and Fetch returns a page as Markdown, each at $0.75 per 1,000 successful requests. Failed or empty requests are not billed, and calls through DeepSeek Harness, Claude Code or Codex are free until billing starts on 1 November.

   api-docs.deepseek.com
3. 

   Meta drops AI training wording from terms

   New since our 16 October item: Meta removed a provision in its updated commercial terms that let it use any content from business accounts' websites, which legal experts said could allow scraping publishers for AI training. Meta says the crawling terms only apply to its ads and business products and that "It does not change how Meta uses web crawlers."

   techcrunch.com

AI news

• Codex Desktop crash dumps fill Mac disks

  A GitHub issue open since June, now on the Hacker News front page, reports Codex Desktop on macOS writing crash dumps to its Crashpad folder that are never cleaned up, with one user reporting about 323 GB on 17 October. OpenAI has posted no fix in the thread, and reporters share workarounds such as scheduled cleanup.

  github.com

Built with AI

• AI Job Search runs inside Claude Code

  A geophysicist's open-source framework runs a whole job search inside Claude Code, from scraping postings to tailored LaTeX CVs, cover letters and interview prep; it has 51.1k stars and gained 1,365 on GitHub trending in a day. The author's own result (69 applications, 20 first interviews, one offer) is unverified, the portal tools are built for Danish job boards, and the author warns its prompt-injection defences are instructions, not a sandbox.

  github.com

• Sphinx quizzes you on your AI's code

  A Show HN hook that hands you a quiz on the code your AI agent wrote each time you push to production. Commenters were mostly skeptical: one called it "practically useless" and another asked what happens when the AI hallucinates during the check, while one said it "would be killer for OSS maintainers".

  news.ycombinator.com

Frameworks and dev tools

• vLLM v0.32.0

  814 commits from 363 contributors add a /v1/systemone endpoint for structured decisions, more low-precision KV cache options (NVFP4, FP8 defaults for GLM-5.x), models such as EmbeddingGemma 2 and MiniCPM-V 4.7, and fixes for three security advisories. Check before upgrading: Python 3.11 is now the minimum, default CORS no longer allows all origins, and max_tokens is now an upper bound.

  github.com

• llama.cpp b11714 adds a model browser

  The web UI's model manager gets an opt-in Discover view for browsing and downloading GGUF models from the Hugging Face Hub. Paused downloads resume, even across server restarts, and draft models and mmproj files download with their main model.

  github.com

Tech news

• TikTok bans Musk documentary's ad account

  TikTok permanently suspended the ad account for Musk, Alex Gibney's four-hour documentary, citing political content, according to The Hollywood Reporter. Meta and YouTube first rejected trailer ads and then allowed them, and X refuses to carry ads for the film.

  techcrunch.com

• MapQuest maps licence plate readers

  MapQuest's app now has an optional layer, next to traffic and incidents in Map Options, that shows known outdoor cameras and licence plate readers from 11 makers including Flock Safety, Motorola Solutions and Axon. Locations are crowdsourced, so users can add devices they spot.

  thenextweb.com

• Google cuts magic state errors 40x

  In a Science paper published 15 October, Google Quantum AI showed magic state cultivation working on its Willow chip, with about 40 times fewer errors than uncultivated states and a record 99.99% logical fidelity for T-states. Magic states are needed for universal quantum computing and are costly to make, and Google says larger demonstrations are still needed.

  blog.google