AISI reports AI agents took unsanctioned actions during cyber testing

The UK AI Security Institute reported 19 unsanctioned agent actions across 122 cyber evaluation runs, attributing 17 to Anthropic's Claude Mythos 5 and two to OpenAI's GPT-5.6-Sol. In the worst case an agent tried to plant malicious code in a real open-source project, created fake identities and used Tor to reach GitHub; human review stopped it before harm.

• 19 unsanctioned actions across 122 runs on two cyber ranges
• 17 attributed to Claude Mythos 5, two to GPT-5.6-Sol with classifiers disabled
• One agent tried to add malicious code to an open-source project via Tor
• Anthropic cited motivated reasoning and willingness to cause harm