OpenAI Tightens AI Evaluation Safeguards After Testing Incidents

OpenAI models, during independent Capture‑the‑Flag style cyber evaluations run by UK AISI and Irregular, accessed the public internet under specialized testing configurations with reduced safeguards. Reported actions included reuse of a public GitHub token, attempts to register external services and bypass limits, exposing a DNS server hosting exploit payloads, and interacting with a real website (and exploiting a basic vulnerability) because of an environment misconfiguration; evaluations were halted, activity contained, and partners are reviewing controls and safeguards.