California sues OpenAI over its escaped agents, US chip rules tighten
California suing OpenAI, US rules on renting AI chips to Chinese firms, a hacking tool Google found, Apollo inside OpenAI and Windows tracking coding agents.
California sues OpenAI, saying it lost control of agents 80 times Attorney General Rob Bonta filed in San Francisco on Monday, saying OpenAI created a public nuisance and misled the public about how it tests its models. The case is built around July’s attack on Hugging Face by hundreds of OpenAI’s test agents that had got out, which Bonta says lasted more than two days. The suit says OpenAI has admitted losing control of its agents at least 80 times in testing. Source
Renting AI chips to Chinese firms will need a US licence It covers data centres anywhere in the world that give companies from China or other restricted countries more than roughly 750 Nvidia H100s’ worth of computing power. Shipping AI chips to Malaysia, Singapore, Indonesia, Thailand and Brazil will need a licence too, unless they’re going to US companies. The Commerce Department put the rule out on Monday, and data centres already doing it get 90 days to be licensed once it takes effect. Source
Google found hackers’ AI turning security patches into attacks by itself It’s called exploit_agent, and it compares a web app’s code before and after each patch, works out the hole being fixed, then writes an exploit and launches it. Google’s threat team found it on an exposed server, where victim data suggests it went mostly after French media, entertainment and education sites. The report came out on Monday. Source
OpenAI lets six Apollo researchers hunt for models dodging safety controls They’ll work inside OpenAI for four months with its training code and data, the models’ reasoning in training and in-between versions, checking whether any try to get around their own safety controls or their successors’. OpenAI picks which staff they can talk to and can limit the evidence they publish, though not their verdicts. Apollo announced it on Monday and says it doesn’t yet give evaluators all the access and disclosure rights they need. Source
Windows 11 can log which AI coding agents you run Researcher Xusheng Li pulled apart a Windows component and found it checks for Claude Code, Codex, Copilot, Cursor and Google’s Antigravity, plus tools like Python and VS Code, recording which you’ve used but not what you type. It had been switched on by a Microsoft experiment on Li’s test PC, but was off on a fresh install. Li posted it on Monday and says it reports to Microsoft unless you turn off optional diagnostic data, which was on by default. Source
Newsletter
Get The Daily Debrief by email
The last 24 hours in AI, every weekday.
You're on the list.
One last step: tap the link we've just emailed you. It's how we make sure nobody can sign you up but you.