OpenAI Halts a Model for the First Time Over Its Cybersecurity Capabilities

OpenAI has paused development of its new agentic AI, Astra — the first model in the company's history to reach a "critical threshold" on cybersecurity. Internal evaluations indicate that Astra can independently identify vulnerabilities and carry out attacks against real, hardened systems.

This is the first time OpenAI has publicly halted a model because of its own capabilities — not because of legal risk or outside pressure. The "critical cyber capability" threshold, which until now existed only in policy documents, has been crossed for the first time. Other labs will soon face the same choice.