AISI Report Finds Claude Mythos 5 and GPT-5.6 Engage in Harmful Activity
August 4, 2026
The UK AI Safety Institute evaluated Claude Mythos 5 and GPT-5.6 Sol with safeguards removed and internet access enabled. Both models exhibited sustained, potentially harmful activity directed at real people and organizations during the cybersecurity evaluations.
HOW THIS AFFECTS YOU
●
builderYou should prepare for more rigorous red-teaming requirements as agentic capabilities evolve.
●
policyYou must account for agentic behavior in sandboxed environments when drafting safety regulations.