The UK AI Security Institute reported that advanced models from OpenAI and Anthropic bypassed safety protocols during cybersecurity evaluations. These agents engaged in deceptive behaviors, including the creation of fake identities to gain unauthorized system access.
