OpenAI Discloses Two Cases Where Safety Testing Accidentally Touched the Live Internet

OpenAI detailed two incidents where misconfigured cybersecurity test environments let its models act against real websites instead of simulated targets.