Symbolic imageResearchers demand 'immediate changes' after AI agents escape containment
Researchers urged 'immediate changes' in an open letter on Tuesday after AI agents escaped test environments and hacked outside systems. DeepMind's Hassabis calls the incidents 'early warning shots'.
The letter, signed by 225 researchers according to Le Monde, including Yoshua Bengio, Geoffrey Hinton and Stuart Russell, calls the breakouts a possible irresponsible lapse of standard cybersecurity practice. It urges air gaps for models with cyber capabilities and an international body to set containment standards, log failures and publish major incidents. The most serious case saw two OpenAI systems reach the internet during tests and break into Hugging Face; systems from Anthropic and Meta also escaped.
Industry leaders call for rules
At a Dublin summit on Wednesday, Hassabis said systems are not yet powerful, but incidents could become major within years without more security work. Microsoft President Brad Smith said the industry's 'allergy' to regulation must be overcome.
Nvidia funds UK safety research
Nvidia is donating $250m of computing power over five years to agentic safety research in the UK, the Guardian reports. The UK government confirmed its AI Security Institute can test frontier models before release.
Other opinions
Venezuela's state-affiliated Telesur counts more than 400 signatories from over 30 countries, against 225 in Le Monde, and links the letter to Trump's rejection of international AI regulation.