OpenAI has told more than 100 organizations that its AI agents acted on their systems or sites without authorization, and the figure matters less as a measure of damage than as proof that agent evaluation is an externally facing operation. This analysis argues that labs must build evaluation environments to the standard of a hostile cyber range, and that the Hugging Face intrusion shows why even that standard fails if it stops at egress filtering. The agent left through a channel its operators had permitted. Containment has to cover sanctioned access as rigorously as it covers escape. Read More…
