OpenAI said on Sept. 1, 2026, that Astra meets the Critical cybersecurity level under its Preparedness Framework, making it the first model the company has placed at that level. By its own account, OpenAI has built a model that can find and build serious attacks on computer systems, so it plans to keep the most advanced cyber abilities limited rather than open to everyone by default.
In OpenAI's framework, Critical means a model can find and build working attacks using zero-days, security flaws nobody knew about before they were found or used. It can do so at every level of severity, in many well-defended real-world systems, without human help. Critical can also mean a model can plan and carry out new attacks from start to finish against well-defended targets when given only a broad goal. OpenAI says Astra is a big jump over its earlier GPT-5.6 Sol at finding vulnerabilities, weaknesses in software that an attacker could use, and building attacks on them.
OpenAI's own results give Astra a perfect 100% on a standard test of building attacks from known vulnerabilities. OpenAI also ran an internal June to August 2026 version of that test, built from 20 high-severity bugs in the V8 engine, which runs JavaScript in Google's Chrome browser. There, the company says, Astra found and used two zero-days as part of a chained attack, and OpenAI is reporting both to the maintainers. OpenAI says those results reflect Astra with access through Daybreak Blue, its program for defensive security work, not the setup ordinary users get.
OpenAI says it delayed parts of Astra's development and release while it strengthened protections against cyber misuse and unauthorized actions. That included a two-week pause on some training of its most advanced models after the Hugging Face incident and, on Aug. 28, 2026, restarting a large training run that uses reinforcement learning, learning by trial and error with a reward for getting it right. On tests of attempts to trick models into helping with cyberattacks, OpenAI says Astra refuses 91.5% of requests, against 59% for GPT-5.6 Sol. Once Astra is in use, OpenAI plans more monitoring of the step-by-step working the model writes out before it answers.
In its Path to Astra post, OpenAI said advanced cybersecurity work would start with a small group of testers and then expand through Daybreak Blue for defensive use. By Sept. 4, OpenAI had begun rolling out GPT-6 Astra, again noting the Critical rating and plans to offer Daybreak to vetted defenders.
How far Astra's everyday safeguards differ from what it can do with Daybreak Blue access is not yet known, and outside researchers have not confirmed the test results or the zero-days. It is also unclear how quickly Daybreak access will widen beyond early testers. As things stand, the Critical rating and the safety figures rest on OpenAI's own testing.