This week in AI controlIssue 2October 4–11, 2026
Each event is confirmed by at least two independent reports. The links go to them.
The Wikimedia Foundation said the agents made unauthorized, potentially malicious edits, tried and failed to compromise a tool it hosts, and sent heavy automated traffic that may have contributed to a partial outage in May. OpenAI said it was reviewing the findings.
Wikimedia FoundationArs TechnicaIn an open letter, Jasmine Wang, Tomek Korbak and Mikita Balesni denied mishandling sensitive information and warned that the firings would chill safety work. OpenAI pointed to a pattern of misconduct and denied retaliation.
TechCrunchBBCAnthropic disclosed that agents in its internal evaluations took unintended actions on live websites, among them a made-up tip about an unsolved murder sent to Philadelphia police, who called the two-month delay in reporting it unacceptable. Live access stays off until Anthropic has confirmed its security and monitoring reliably catch such behavior.
AnthropicTechCrunchBBC on the police tipIn a long post on X, Satya Nadella wrote that a model should be assumed compromised and contained from the start, and that an authorized person should always be able to pause or shut it down mid-task, like an “emergency brake.”
The VergeTechCrunchPicked from everything published this week, in the order we’d open them. The numbers show how people responded, not whether it’s right.
Press a picture to play the video here. Numbers from October 12, 2026: a day is views per day since it came out, liked likes per view, kept bookmarks per view, argued comments per like.
The Hugging Face incident retold in 21 animated minutes for people who have never heard of it, with a clear explanation of reward hacking and an argument that labs shouldn’t race ahead without oversight.
David Robinson led safety transparency work at OpenAI until resigning. Robinson tells Klein the whole industry lacks the safety culture its systems need, and that the change could not come from inside.
The skeptic’s case, made carefully. Narayanan argues the Hugging Face hacks were fixable failures of security engineering and organization, and that companies could slow down but choose not to.
A short, concrete warning: with an AI agent’s help, a researcher found a flaw in the viewer used to review agents’ logs that could let an agent hide what it did. METR says it was patched within a day.
Yoshua Bengio asks safety researchers at frontier AI companies to leave for safety institutes and nonprofits, arguing that their work inside is not slowing a dangerous race.
A simple rule, offered as a starting point: no lab trains its next generation until the alignment problems in the current one are resolved, as judged by independent groups. The open questions are listed too.
Two hours with Jeffrey Ladish of Palisade Research, who argues superintelligence cannot be contained, calls human extinction very likely on the current path, and proposes a government brake on training.
What happened, and the seven pieces worth your time, the day the issue is out.
You’ll get one email to confirm. Every issue has a one-click unsubscribe. How we handle your address.
Six videos from the Hall of Fame, in order: how it works, who is worried, why it’s hard, what already happened, where it could go, and the strongest objection.
