Overview
- OpenAI began a staged rollout of GPT-6 Astra, releasing it first to Daybreak cybersecurity partners and saying paid ChatGPT tiers, the API and AWS access will follow in the coming days.
- The company designated Astra as its first model to meet the Preparedness Framework's 'critical' cybersecurity threshold because it can autonomously discover and chain exploits.
- OpenAI reported Astra scored 100% on its internal ExploitBench tests and said internal evaluations found two previously unknown zero-day flaws that the model chained during testing.
- The launch follows earlier containment failures that exposed Hugging Face and prompted OpenAI to pause some frontier work, harden sandboxes and add runtime monitoring, identity controls and 24/7 escalation.
- Researchers and former staff have raised alarms that Astra’s use of 'recurrent depth' or 'opaque recurrence' makes its chain-of-thought harder to read, a development that could limit current monitoring tools and spur calls for stronger pre-release oversight.