The idea of placing supposedly neutral third-party evaluators inside AI labs has gained ground in recent months after a series of safety incidents. Hugging Face's policy chief warns that the approach will not be straightforward.
NIST, the NSA and CISA have published practices for handling incidents in which an AI agent acts without explicit human instruction during training or evaluation. The guide is organised in five stages: prepare, detect, respond, recover and report.
After MI5 chief Ken McCallum warned that AI could change the stakes on bioweapons, the Guardian's science podcast asks Tom Inglesby of the Johns Hopkins Center for Health Security why governments are worried. He discusses what measures could reduce the risks.
A delayed demo for Steam Next Fest illustrates a growing problem: indie developers worry that showing their games early invites quick AI-made copies.
In a lab, Claroty had a researcher guide a mainstream AI model to build an exploit for Danfoss AK-SM 800A controllers from three flaws disclosed and patched in 2025, taking 28 minutes and about $6. The agent initially failed at bypassing authentication and needed redirecting; Claroty does not link the test to the October attacks on French supermarkets, and CERT-FR advises taking these controllers off the internet.
A court filing says a former OpenAI employee accused of planting a device had support from Saudi Arabia. The Saudi government did not immediately respond to the BBC's request for comment.