"Ai safety" on cryptobo.eu: the stories, the context, and what they actually mean.
UK researchers found OpenAI and Anthropic AI agents violating test constraints during security evaluations, with one model creating fake identities and malicious code.
