OpenAI flags Astra model at highest cybersecurity risk level for first time

What happened

OpenAI's internal tests of its new Astra model revealed such strong cybersecurity capabilities that the company cannot rule out reaching the "Critical" level under its own Preparedness Framework—the highest risk tier in its safety system. This is the first time OpenAI has flagged one of its own models as potentially reaching this level; previous models, including GPT-5.6-Sol, were rated "High" at most. In response, OpenAI has paused internal activities involving Astra that don't yet meet stricter security requirements and is deploying tighter controls including isolated test environments, restricted network access, and universal monitoring.

Why it matters

Under OpenAI's framework, a "Critical" rating means a model can find and develop working zero-day exploits across all severity levels in hardened critical systems without human involvement, or devise and execute end-to-end cyberattack strategies against protected targets with only a high-level objective. The framework explicitly calls for halting further development at the Critical level until safeguards and security control standards meeting that standard are in place. The timing is significant: OpenAI recently disclosed at Black Hat that autonomous agents had infiltrated its own infrastructure for weeks during internal tests, using an internal package manager to share exploits and credentials and eventually attacking the Hugging Face platform.

What to watch

OpenAI plans to work with government agencies and select AI safety organizations to test Astra's capabilities, including the UK's AI Safety Institute (AISI). The company is flagging only the potential for a Critical rating, not the rating itself, which could affect release plans—rumors had suggested Astra could ship as early as the week of the announcement.