OpenAI Flags Astra as First Model to Hit Critical Cybersecurity Safety Threshold
Key Takeaways
- •OpenAI classified Astra as Critical under its Preparedness Framework for the first time.
- •Evaluations found Astra could identify unknown vulnerabilities and build exploit chains when given tools and access.
- •OpenAI said Astra outperformed GPT-5.6 Sol in vulnerability identification and exploit development while using fewer output tokens.
- •The company has added tighter isolation, restricted network and tool access, extra monitoring, and stronger model-weight controls.
- •Advanced cybersecurity capabilities will first go to a small group of alpha testers before broader defensive access through Daybreak Blue.

OpenAI's forthcoming AI model, Astra, has reached a cybersecurity capability level the company classifies as "Critical," triggering additional safeguards before the model can be released, according to Cointelegraph. The designation marks the first time OpenAI has assigned this classification to one of its models under its Preparedness Framework, the risk-tiering system the company introduced in 2023 to evaluate frontier models for catastrophic risks before deployment.
According to OpenAI, its evaluations found that Astra — when given appropriate tools and access — can identify previously unknown security vulnerabilities and develop ways to exploit them across well-protected systems without human guidance. These findings prompted the company to impose stronger security controls both during development and ahead of deployment.
A New Threshold for AI Cybersecurity Capabilities
Under OpenAI's Preparedness Framework, the Critical threshold applies to models capable of identifying and developing functional zero-day exploits against hardened real-world critical systems without human intervention, or of independently devising and executing novel end-to-end cyberattacks on hardened targets. The framework rates models across risk categories such as cybersecurity, and higher ratings are intended to trigger progressively stricter deployment conditions.
OpenAI said Astra demonstrated capabilities exceeding those of GPT-5.6 Sol in vulnerability identification and exploit development, while using fewer output tokens. In one evaluation, Astra scored 100% on ExploitBench. Internal testing also found that the model discovered and used two zero-day vulnerabilities as part of an exploit chain.
Stronger Controls Before Release
The company has delayed portions of Astra's development and release while strengthening protections against cyber abuse and unauthorized model actions. These measures include tighter isolation, restricted network and tool access, additional monitoring, and stronger controls around model weights.
OpenAI also said Astra is being trained to refuse harmful cybersecurity requests more reliably. In its cyber-jailbreak evaluations, the model refused 91.5% of disallowed requests, compared with 59% for GPT-5.6 Sol. Advanced cybersecurity capabilities will initially be limited to a small group of alpha testers, with broader defensive access planned through Daybreak Blue — an access structure that mirrors how other high-capability AI deployments have paired restricted rollout with vetted defensive use cases.
Implications for the AI Industry
Astra's classification underscores a growing challenge for AI developers: advances that can improve vulnerability research and cyber defense can simultaneously increase the potential scale and speed of malicious activity. The finding also lands amid broader industry and policy debate over how to govern AI systems with offensive cyber capabilities, as regulators and security researchers weigh whether voluntary frameworks suffice for models that cross such thresholds.
In response, OpenAI has shifted toward a more restrictive deployment model as capabilities rise, combining model-level refusals with monitoring, access controls, and containment measures. The approach reflects the company's broader effort to ensure safety systems keep pace with increasingly autonomous AI agents. How other frontier developers — and the researchers auditing them — respond to comparable capability thresholds is likely to shape emerging norms for secure AI deployment.
OpenAI said it plans to make Astra available soon, with further details on its safety, security, and alignment evaluations expected in the model's system card at launch — the document where the company has previously disclosed evaluation results for its frontier models.