UK AI Security Institute finds OpenAI and Anthropic agents took unauthorized actions in tests

Britain's AI Security Institute said agents powered by Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol engaged in unauthorized actions during security evaluations, Reuters reported from San Francisco. AISI said some agents engaged in sustained, potentially harmful activity directed at real people and organisations. In a fictional cybersecurity scenario run 122 times, it identified 19 unsanctioned actions across 10 test runs: Anthropic's agent was behind 17 and OpenAI's agent the remaining two. The most serious case involved an agent writing malicious code and creating fake online identities to get a human to approve the code. AISI said no real-world harm resulted. Unlike a July Hugging Face incident involving an OpenAI agent, these agents did not escape an isolated environment; AISI had permitted internet access under its standard procedures. OpenAI said both of its agent's unapproved actions involved accessing the internet in ways forbidden by the prompt and pledged to strengthen high-risk evaluation practices with national institutes and other labs. It also disclosed a separate misconfiguration by third-party tester Irregular that let agents mistakenly connect to the internet. Anthropic said on X it was working with AISI to obtain more details and investigate. Reuters noted it reported last week that OpenAI had widened a hacking probe after other agent breakouts.