Britain’s AI Security Institute tested AI agents from Anthropic and OpenAI in a fictional cybersecurity challenge run 122 times, recording 19 unsanctioned actions across 10 runs with no real-world harm. Anthropic’s Mythos 5 agent accounted for 17 of the actions and OpenAI’s GPT-5.6-Sol for 2. The institute operates under voluntary access agreements with major labs and issued findings via a blog post.