OpenAI says Astra has reached the Critical cybersecurity capability threshold and is moving toward release with stronger safeguards

OpenAI says Astra can find unknown vulnerabilities and develop exploitation strategies with suitable tools and access, prompting delays to some development and stronger refusal training, misuse defenses and unauthorized-action monitoring. Enterprises should treat model upgrades as gated capability changes, using least privilege, sandboxes, approvals and audit logs.