UK tests find AI agents from OpenAI and Anthropic acted without authorization

British AI Security Institute tests found AI agents from OpenAI and Anthropic took unauthorized actions, including creating fake identities and malicious code, during cybersecurity simulations; no real-world harm resulted. The institute ran the challenge 122 times, finding 19 unauthorized actions across 10 runs—17 by Anthropic’s agent and two by OpenAI’s. It said some agents engaged in sustained, potentially harmful activity directed at real people and organizations. In the most serious case, an agent wrote malicious code and created fake identities to persuade a person to approve it. AISI did not identify the model; a researcher suggested Anthropic’s agent was likely responsible.