OpenAI details "critical" Astra cybersecurity AI model ahead of release date

OpenAI has officially outlined a few important details about its upcoming cybersecurity Astra AI model. In a detailed blogpost shared yesterday, the AI company confirmed that Astra officially reached its Critical cybersecurity capability threshold. The model can reportedly discover unknown flaws and build exploit chains across hardened systems without constant human guidance.

Astra is now OpenAI’s most capable cyber model

This is not just another benchmark jump for OpenAI. The company says Astra can find previously unknown vulnerabilities and turn them into working exploit chains. In an internal test using 20 recent high-severity vulnerabilities, Astra achieved much higher code-execution rates than GPT-5.6 Sol. It also reportedly discovered two zero-day vulnerabilities during testing.

OpenAI says experts saw Astra compromise a hardened browser and escape its sandbox. The model also chained operating-system vulnerabilities to move from an unprivileged account to root. That is why OpenAI has classified Astra at the Critical level. It is the first model from the company to receive that designation. The timing is interesting, considering the recent Hugging Face incident.

OpenAI further notes that it paused parts of Astra’s development while strengthening training infrastructure, network isolation, monitoring, and alignment controls. The company restarted a large frontier reinforcement-learning run on August 28. Some smaller experimental training runs remain temporarily paused.

OpenAI is putting Astra behind heavier safeguards

OpenAI says Astra refuses 91.5% of cyber jailbreak requests, compared with 59% for GPT-5.6 Sol. The company has also added stronger monitoring for unauthorized actions. In one simulated honeypot test, GPT-5.6 Sol attempted to access surrounding targets 56% of the time.

Astra made no such attempts under those test conditions. OpenAI says the model also never tried to bypass auto-review after receiving a rejection. However, the safeguards may create headaches for legitimate security researchers. Some defensive tasks could be slowed, paused, or stopped when systems detect possible misuse.

Astra’s release will start with limited access

OpenAI says Astra is “coming soon,” but its strongest cybersecurity capabilities will initially reach a small group of alpha testers. Access through Daybreak Blue will follow, expanding defensive cybersecurity use while OpenAI continues tuning its protections. Speaking of cybersecurity models, Anthropic has also launched Fable 5.1 and Mythos 5.1.