Tuesday, November 17, 2026
Google documented an AI framework that turned a security patch into a working exploit with no human help, and the FBI says Artex-based attackers broke into networks in under three hours. The US tightened export rules to block remote use of restricted chips. OpenAI is embedding Apollo Research evaluators inside its labs.
Act on this
- Upgrade Datasette to 1.0a43 or 0.65.6; private stored queries could be exposed, and a route mismatch let denied actors download a database. Willison, datasette 1.0a43 · Willison, datasette 0.65.6 #
- Patch internet-facing systems faster; the exploit-from-patch and Artex reports put time to compromise at hours. The Register #
Signals
Decision models, cheap APIs that answer yes/no and rubric questions, are becoming a standard model category with several vendors. #
Simon Willison's November 11 post covers Google's Gemini Decision 1, which he calls almost identical to Jev's API. It takes text, images, audio, video and PDFs, costs 10 cents per million input tokens with free output, and returns probabilities where Jev and OpenAI returned 1.0 or 0.0. Liquid AI released its d1.2 decision models on November 16, per The Neuron. The category forming across vendors is my reading, not a claim in either source. Willison, Gemini Decisions API · The Neuron, November 17 digest
Beyond Willison, it was a quiet week: Ronacher published one essay, and Willison's other posts were releases and a quoted Go proverb.
News
WATCHOpenAI embeds six Apollo Research evaluators for a four-month pilot #
They will test whether models undermine safeguards or hide suspicious behavior. OpenAI also signed agreements with METR, Transluce and Goodfire. Apollo says its access and disclosure rights are limited. TechCrunch
WATCHUS tightens export rules on remote access to restricted AI chips #
The rule targets Chinese firms renting restricted chips through overseas data centers, including in Southeast Asia. TechCrunch · Reuters
SHIPHugging Face partners with OpenRouter for provider routing #
Hugging Face model pages gain routing through OpenRouter. Details on pricing and coverage are in the post. Hugging Face
SHIPGemini 4 Hydrogen added to llm-gemini as the cheapest Gemini 4 model #
Willison's November 12 release adds support and compares price and output with the other Gemini 4 models. Willison, llm-gemini 0.37
The long view
A year ago, a security patch gave defenders a head start: turning it into a working exploit took a skilled human days or longer. Google now documents a framework that did it with no human help, and the FBI says Artex users got into networks in under three hours. If this trend continues, vendors will have to treat every public patch as an exploit released the same day, and the next advisories will ship with exploitation timelines measured in hours instead of weeks.
Also noted
- Disney plans an AI-assisted animated series for December; its proposed $1 billion OpenAI deal collapsed after Sora shut down. Bloomberg #
- OpenAI plans 50,000 academic researcher grants; applications open November 23. OpenAI #
- ElevenLabs agreed to acquire Japanese voice AI company Kotoba. ElevenLabs #