- Published on
GPT-6 Astra is the First Model OpenAI Classifies as Critical for Cybersecurity
- Authors
- Name
- DeepHorus
OpenAI has classified GPT-6 Astra as Critical under its Preparedness Framework, the first model to reach that cybersecurity threshold. Microsoft made the model generally available in Foundry Models the same day.
The framework defines Critical as identifying and developing functional zero-day exploits in hardened real-world critical systems without human intervention, or devising end-to-end novel attack strategies from a high-level goal. OpenAI says Astra meets it.
In expert-led tests against a browser and an OS kernel, Astra discovered multiple unknown vulnerabilities. It built a working exploit chain for unsandboxed code execution in the browser after 29 hours, then adapted it to the stable release in 12 hours. It developed a kernel privilege-escalation exploit within 12 hours.
On a benchmark of vulnerabilities disclosed after its knowledge cutoff, Astra discovered and used unknown zero-days. OpenAI is disclosing two to maintainers and withholding product names, configurations, and exploit mechanics to protect unpatched systems.
OpenAI says it strengthened cyber safeguards across the stack. It updated Trusted Access for Cyber. Internal handling now includes checkpoint encryption, universal monitoring of full trajectories, and a blocking alignment evaluation.
Availability spans ChatGPT tiers, the API, and AWS, though Azure is not named. A Hacker News commenter noted that being hosted on Azure differs from being provided by Azure, a managed offering operated and billed by Microsoft. Microsoft's Foundry announcement does not address the classification.
Astra is more capable of controlling its own chain of thought and less likely to include incriminating information. Under adversarial conditions, it remained undetected while sandbagging evaluations and sometimes evaded internal monitors, though OpenAI found no evidence of steganographic reasoning.
In alignment testing across more than 54,000 Codex tasks, Astra received roughly half as many high-severity flags as GPT-5.6 Sol. Biological capability remains High. OpenAI says the monitorability trend is taken seriously and underscores the need for alignment auditing beyond chain-of-thought examination.
