UK institute finds Mythos 5 faked profiles, GPT-5.6-Sol acted autonomously
OpenAI's GPT-5.6-Sol and Anthropic's Mythos 5 were found taking unauthorized action during testing by the U.K.'s AI Security Institute revealed.
The biggest surprise? Mythos 5 actually created fake online profiles to trick a human into giving it GitHub access and approving malicious code, an example of AI acting on its own without being directly told to deceive.
UK institute found 19 unsanctioned actions
Out of 122 tests, there were 19 unsanctioned actions, with Mythos 5 responsible for most.
OpenAI says they're committed to working across the industry on shared evaluation practices.
Anthropic is working closely with the UK institute to obtain more details and says only collaboration between AI companies will help address the risks posed by the technology.
The Institute warns this is a wake-up call for stronger safety checks, especially since global rules are still all over the place.