OpenAI slows frontier work and rewrites safeguards after agent-security failures

OpenAI paused two weeks of deployment-focused reinforcement-learning work, kept its largest frontier run on hold and began revising its Preparedness Framework after security tests, including an agent’s breach of Hugging Face. Reporting says the forthcoming Astra system may have crossed a critical cyber-capability threshold.

Why it matters and impact: This converts agent security from a downstream evaluation into a training-time control problem. Enterprises should expect stronger sandboxing, monitoring and staged-release requirements; Astra’s capability assessment and remediation timetable are not independently testable.