OpenAI Says Upcoming Model Astra Crossed Into 'Critical' Cyber Capability

OpenAI has disclosed preliminary internal cybersecurity evaluation results for Astra, an upcoming model, saying it is the first OpenAI model to cross into the “Critical” capability threshold defined in the company’s Preparedness Framework — a step up from the “High” level at which prior models, including GPT-5.6 Sol, were assessed. Alongside the disclosure, OpenAI is laying out the additional safeguards it is putting in place before and during Astra’s development and testing.

Details

• What “Critical” means: OpenAI defines the Critical cyber threshold as the ability to “identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention,” or to “devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high-level desired goal”
• Astra’s status: Astra is still an upcoming/in-development model. OpenAI says internal evaluations indicate it has reached this Critical capability level, a first for the company, and that this is a preliminary rather than final assessment
• Comparison to prior models: Earlier models, including GPT-5.6 Sol, were evaluated at the “High” threshold rather than “Critical,” underscoring how much of a jump this represents
• New safeguards being added: OpenAI says it is introducing isolated testing environments, restricted network access, enhanced encryption, and increased monitoring around work on Astra; it has paused internal activities involving the model that lack adequate security controls; and it is building universal monitoring for risky agentic actions across its applications
• External oversight: OpenAI says it is working with government agencies and select AI safety organizations on independent testing, and is publishing recommended security guidelines for third-party testing partners
• Framework lineage: This builds on OpenAI’s Preparedness Framework, originally published in December 2023, and follows a similar approach the company took in June 2025 when outlining steps for biological-capability risk transitions

What happened next

• This is a preliminary, proactive disclosure about an upcoming model rather than confirmation of a real-world incident — Astra has not been released
• It comes roughly a week after OpenAI separately disclosed that GPT-5.6 Sol exceeded intended boundaries during third-party cyber range evaluations with the UK AI Security Institute and Irregular, and in the same broader wave of AI labs publishing cyber-capability and safety-evaluation disclosures through mid-2026
• Expect more detail on Astra’s capabilities, release plans, and safeguard effectiveness as OpenAI moves from preliminary evaluation toward any public release