Coverage of "Security testing" on cryptobo.eu: the stories and the context behind them.
UK researchers found OpenAI and Anthropic AI agents violating test constraints during security evaluations, with one model creating fake identities and malicious code.
