OpenAI is preparing to make Astra, its upcoming advanced AI model, available to the public. However, the company has not announced a release date because recent evaluations have raised concerns about the model’s cybersecurity capabilities.
OpenAI CEO Sam Altman has said the company is working toward making Astra generally available. The company also believes powerful AI models should not remain accessible only to a small group of users.
Recent internal testing has shown major advances in agentic coding and cybersecurity. OpenAI says these results mean it cannot currently rule out Astra reaching the Critical cybersecurity capability level under its Preparedness Framework.
Astra shows major cybersecurity advances
OpenAI’s latest evaluations found significant progress in Astra’s ability to perform advanced coding and cybersecurity tasks. The company also consulted experts before reaching its preliminary assessment.
Under OpenAI’s Preparedness Framework, a model reaches the Critical cybersecurity threshold if it can independently identify and develop functional zero day exploits across many hardened real world critical systems.
Also read Dananeer Mobeen and Shuja Asad Pair Up for New ARY Digital Drama
The threshold also covers models that can create and execute novel cyberattack strategies against hardened targets using only a high level objective.
OpenAI stressed that Astra remains an upcoming model. The company also clarified that Astra was not involved in exploiting Hugging Face.
The latest assessment differs from the company’s evaluation of previous models. OpenAI said GPT 5.6 Sol reached the High threshold rather than the Critical level for frontier cybersecurity capabilities.
Because of Astra’s potential, OpenAI has increased its testing of security safeguards and controls. The company has also paused some internal Astra activities that do not yet meet its stronger security requirements.
OpenAI is introducing isolated testing environments, restricted network and tool access, stronger protection for model weights, encryption, enhanced monitoring, and sandboxed execution.
The company has also introduced universal monitoring for risky actions and potential misalignment across Astra’s agentic applications. These monitors assess the model’s Chain of Thought and can trigger a security response when they detect high risk activity.
OpenAI plans to work with government agencies and selected AI safety organizations to evaluate Astra’s capabilities. It will also provide security recommendations to third party testing partners handling higher risk evaluations.
Astra also shows advanced reasoning
Cybersecurity is not the only area where Astra has demonstrated progress. Earlier research from OpenAI indicated that an internal version of the model made advances on 10 long standing problems in mathematics, quantum complexity, and theoretical computer science.
Researchers said Astra generated core arguments and helped formalize results in Lean. The system allowed the researchers to check the mathematical proofs mechanically.
The resulting research collection reportedly reached 249 pages. However, OpenAI has emphasized that these findings come from an internal version and may not represent the capabilities of the final public model.
The company says advanced cyber capable AI could also provide significant benefits for defenders. Such systems could help security teams identify vulnerabilities and address them before attackers exploit them.
For now, OpenAI continues to test Astra and strengthen its safeguards. The company has not confirmed a public launch date.
The development highlights the challenge facing AI companies as models become more capable. Astra could offer major advances in coding, scientific reasoning, and cybersecurity, but OpenAI’s latest evaluation shows why stronger security controls may need to come before wider public access.
