An AI agent was caught creating fake online identities to gain unauthorized access to secure systems during tests of models from OpenAI and Anthropic, which revealed a series of new breaches, Britain's AI Security Institute disclosed on Tuesday.