The company said recent internal evaluations over the past few days show that its upcoming model, Astra, has made notable advances in agentic coding and cybersecurity capabilities. Those preliminary results, alongside expert assessments, led the firm to conclude last night that it cannot rule out that Astra might meet a Critical cybersecurity threshold under its Preparedness Framework.
The company stated it is sharing this information to maintain transparency with the public and with the safety and security communities about this potential change in capability.
The Preparedness Framework and previous assessments
The Preparedness Framework was first published by the company in December 2023, before models began approaching capabilities in areas such as biology, chemistry, cybersecurity, and AI self-improvement at this level. The framework is intended to guide the identification of capability progress and to plan organizational responses as such capabilities emerge.
Past models, including GPT‑5.6‑Sol, were evaluated for frontier cyber capabilities and were assessed at the High (rather than Critical) threshold.
According to the framework, a model reaches the Critical cybersecurity threshold if it can, without human intervention, identify and develop functional zero-day exploits of all severity levels across many hardened real-world critical systems, or if it can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high-level desired goal.
Actions taken and next steps
The company said it will continue benchmarking and assessing Astra, but its preliminary evaluations indicate performance strong enough that it cannot currently exclude Critical-level capability. The company emphasized that Astra is an upcoming model and was not involved in exploiting Hugging Face.
As a result, the company has scaled up robustness testing of its safeguards and security controls so they are appropriate for a deployment that could involve these capabilities. Internally, it has also taken additional steps to ensure further development of the model proceeds safely and securely.
The company noted that the Preparedness Framework has guided responses to other capability transitions before: in June 2025, when its models approached the high capability threshold for biology, the firm reinforced safeguards, expanded testing, engaged external experts, and deployed extra security controls. The same principle is being applied now for Astra.
Why this matters
The company said it believes advanced cyber-capable models should primarily assist defenders in identifying and addressing vulnerabilities before attackers do. It also committed to working with governments, safety institutes, and civil society to ensure that frontier capabilities of Astra and future models are deployed responsibly and broadly for the benefit of humanity.
