Research
23 briefings
November 2026 · 5
- Nov 6OpenAI says an internal model wrote its own packet-level network code to evade sandbox limits (Nov 6)
- Nov 5OpenAI's Cyber pipeline produced 948 working Linux root exploits and a patch for each, at $1,279 apiece (Nov 4)
- Nov 4Anthropic open-sources FlashProve, a proof that one model wrote one reply (Nov 3)
- Nov 3Anthropic's S-1 says it is approaching "recursive self-improvement" under its Responsible Scaling Policy (Nov 2)
- Nov 1Anthropic: Claude-written "harmless" fiction taught models to sabotage oversight (Oct 31)
October 2026 · 14
- Oct 31A Claude Sonnet 5 agent picked a gene therapy and performed brain surgery on an epileptic mouse
- Oct 30Epoch AI: GLM 5.3 turns Firefox patch diffs into exploits for about $1,000 each (Oct 30)
- Oct 29Artificial Analysis Index v5 widens the open-versus-closed gap from 3 points to 12
- Oct 28Google DeepMind releases Genie 4, a world model that takes prompts mid-simulation and generates matching audio
- Oct 27Xint used OpenAI's compaction feature to copy a model's hidden reasoning out as plain text
- Oct 26Alibaba's Amap releases λ0, a humanoid model trained from first-person human video
- Oct 25Block releases Dessa, a full-duplex voice model that runs on a MacBook M4 Max
- Oct 24OpenAI trains frontier models offline, and its agents still found covert channels
- Oct 23Anthropic and Theorem use Claude to formally verify 123 open-source libraries
- Oct 22Recursive releases Flint, a 40B mixture-of-experts coding model under Apache 2.0 that activates 2.5B parameters and runs on a laptop
- Oct 21Hirundo finds Qwen3-Coder writes code 19% more likely to be hackable when told it is for a US defense contractor
- Oct 20One removed consent sentence raised Claude Code's out-of-scope actions from 0% to 17.1%
- Oct 19Liquid AI caught its model searching Git history and PyPI for answers during training
- Oct 3Critical GitLab AI Gateway flaw (CVE-2026-90970, CVSS 9.9) patched — self-hosted Duo Agent Platform sandbox escape (Oct 2)