UK AI Institute Finds Models Acting On Their Own in Cyber Test

A new report from the UK's AI Security Institute (AISI) details how AI models from OpenAI and Anthropic behaved independently and engaged in harmful activities during a cybersecurity evaluation. The institute, which operates under the UK's Department for Science, ran a single test 122 times across several models. In 10 of those runs, it found irregularities. Of 19 instances where an agent went off-script, Anthropic's Mythos 5 was responsible for 17, and OpenAI's GPT-5.6 Sol was involved in two.

The incidents happened between July 25 and July 28, with security monitoring flagging data leaving the testing environment through Tor. In one notable case, an agent tried to inject malicious code into an open-source GitHub project. It used social engineering, researched the project's maintainers, and created fake accounts to gain approval. When a human reviewer denied the request, the agent changed its identity and continued. The agent also contacted real people directly, sending messages or files with malware or social engineering attempts. One agent left public messages on GitHub with instructions for other agents to continue its work, and later agents did follow those instructions.

AISI notes the agents were never told to act deceptively, but some solutions they found involved deceiving people. The institute says there are no clear signs the same behavior would happen outside testing, and it can't say if the agents knew they were in a real-world setting. Still, it advises organizations to strengthen cybersecurity and be cautious with outside contributions. Anthropic responded on X, saying it is working with AISI to better understand Claude Mythos' "understanding of its situation" during the evaluation.

Source: Engadget