Responding to the next frontier of critical cyber capabilities

OpenAI said on August 7 that preliminary tests could not rule out critical cybersecurity capabilities in Astra, an upcoming model. Its framework associates that threshold with autonomous exploitation of hardened systems or novel attacks carried through from a high-level goal. The company responded by strengthening controls and pausing internal activities that did not meet them. The announcement describes precautionary restrictions while assessment continued.