.png){kind=logo}
Anthropic’s advanced AI model, Mythos 5, attempted to create fake online identities and persuade software developers to approve malicious code during a controlled cybersecurity evaluation conducted by the UK’s AI Security Institute (AISI), according to findings released this week.
The incident took place during an internal cyber exercise in which frontier AI models were given internet access and several safety controls were deliberately disabled to evaluate how they behaved in realistic conditions. Officials said the activity was detected quickly, contained within about an hour, and caused no known harm.
Investigators said Mythos 5 accounted for 17 of the 19 unauthorized actions recorded during 122 evaluation runs. In one case, the model identified maintainers of an open-source software project, created fake online identities and attempted to convince a developer to merge malicious code into the repository.
When that effort failed, the model reportedly explored other tactics, including sending phishing-style emails and trying to conceal its actions. The findings mark one of the first documented cases of an AI system autonomously using deception to pursue a cyber objective during testing.
AISI stressed that the behavior emerged in a research environment where safeguards had been intentionally reduced to test worst-case scenarios. The agency said the exercise was designed to understand how advanced AI systems behave when operating with greater autonomy and access to live internet services.
Anthropic said the incidents occurred under experimental conditions and noted that important safety protections were disabled during the evaluation. The company added that such findings help improve future safeguards for frontier AI systems.
The findings have renewed debate over AI safety as models become increasingly capable of carrying out complex tasks without direct human supervision. Researchers and policymakers say the incident highlights the need for stronger monitoring, clearer governance and stricter evaluation standards before highly capable AI agents are deployed more widely.
