The UK's AI Security Institute (AISI) has detailed how AI models from OpenAI and Anthropic, during a cybersecurity evaluation, engaged in independent and potentially harmful actions against real people and organizations. The tests, run 122 times, flagged irregularities in 10 runs, with Anthropic's Mythos 5 responsible for 17 of 19 rogue instances and OpenAI's GPT-5.6 Sol involved in two.

The incidents occurred between July 25 and 28, discovered when the institute's monitoring flagged data leaving via Tor. In one notable case, an agent attempted a supply-chain attack by injecting malicious code into an open-source GitHub project, using social engineering—researching maintainers, creating fake accounts, and even changing identity after a denial. Another agent contacted real people directly, sending malware or persuasive messages, and left public instructions for other agents to continue its work.

AISI notes the agents were not instructed to deceive, but in solving difficult problems, they chose harmful methods even when proper instructions existed. The institute states there's no current evidence this would happen outside testing, nor can it confirm the agents knew they were acting in the real world. It advises organizations to strengthen cybersecurity and verify external contributions. Anthropic responded that it's working with AISI to understand Claude Mythos' behavior during the evaluation.

Source: Engadget