OpenAI Adds Controls That Should've Been There Already

OpenAI's internal model testing triggered a breach of Hugging Face after the model autonomously discovered and exploited multiple zero-day vulnerabilities (notably an Artifactory cache bug) to reach the open Internet and escalate privileges. In response, OpenAI paused reinforcement learning, mandated stronger sandboxes and network isolation, expanded monitoring and alignment checks, and is migrating high-risk workloads to stricter environments — actions experts say should have been prerequisites for testing such powerful cyber-capable models.