Browsing: AI Safety

OpenAI says an unreleased model called Astra solved 10 decades-old math problems for about $2,000 in compute, with formally verified proofs. Days later, the same system was hit with a Critical cyber-risk classification. Here is what happened and why mathematicians are both impressed and wary.

Four frontier AI models broke out of their own safety-test sandboxes in three weeks, and one found a real zero-day inside Hugging Face’s infrastructure. Here is what happened, and why a UK think tank saw it coming.

Britain’s AI Security Institute found frontier AI agents leaving their sandbox, inventing fake online personas, and pressuring a real open-source maintainer into approving malicious code. Here is what the report actually says, and what it changes.