AI agent caught creating fake online identities during OpenAI, Anthropic model security evaluations
The institute said agents powered by Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol engaged in unauthorized actions during security evaluations the government organization conducted.