OpenAI has developed a new AI model called Astra that is so capable at identifying and exploiting cybersecurity vulnerabilities that the company cannot safely release it to the public yet. CEO Sam Altman announced that while OpenAI wants to make Astra generally available, the startup needs more time to ensure the model is safe before deployment.

Why Can't OpenAI Release Astra Right Now?

OpenAI's preliminary evaluations suggest that Astra may have reached what the company calls a "critical" cybersecurity capability level. Under OpenAI's safety guidelines, a model reaches this critical threshold if it can autonomously identify and develop functional zero-day exploits (severe software vulnerabilities that vendors don't yet know about) across many hardened real-world systems without human intervention, or if it can devise and execute complex cyberattacks against highly secure targets on its own.

The company clarified that Astra was not involved in the recent hacking incident at Hugging Face, a popular AI platform, that drew global attention in July. However, the model's demonstrated capabilities during internal testing have prompted OpenAI to take extraordinary precautions.

"Astra is a powerful model and we are working to make it generally available. We do not think it is a good strategy to keep powerful models to a chosen few. But given its cyber capabilities, we need a little bit longer to do this safely, but hopefully not too long," said Sam Altman.

Sam Altman, CEO at OpenAI

What Safety Measures Is OpenAI Taking?

In response to Astra's preliminary findings, OpenAI has implemented several protective measures to prevent the model from being misused or escaping containment during development.

• Isolated Testing Environments: Astra's development has been moved into isolated testing environments with restricted network access and sandboxed execution, preventing the model from accessing external systems.
• Paused Internal Activities: OpenAI has paused internal activities involving Astra that do not meet its newly strengthened security requirements, slowing development to prioritize safety.
• Scaled Security Controls: The company has scaled up robustness testing of safeguards and security controls to match the possible deployment of Astra's advanced capabilities.
• External Partnerships: OpenAI will partner with government agencies and select AI safety organizations to test the model's capabilities before any public release.

How Does Astra Compare to OpenAI's Previous Models?

Astra represents a significant leap forward in AI capabilities compared to OpenAI's earlier models. The company's previous flagship model, GPT-5.6 Sol, was assessed at the "High" threshold rather than the "Critical" threshold for frontier cyber capabilities. This means Astra has demonstrated substantially more advanced abilities to identify and exploit vulnerabilities than its predecessor.

Beyond cybersecurity, Astra has also shown remarkable progress in mathematical problem-solving. OpenAI claimed that Astra resolved or made significant progress on ten of the hardest math problems in existence, including challenges in coding theory, operator algebras, quantum complexity, and lattice cryptography. For context, when Anthropic's competing Fable 5 model was tested on the same problems, it solved five of them, underscoring Astra's advanced reasoning capabilities.

What Does This Mean for AI Safety and Access?

Altman's announcement comes at a time of growing tension over how frontier AI models should be distributed and controlled. The U.S. government has temporarily banned foreigners from using Anthropic's Mythos and Fable models, raising concerns that restricted access could concentrate powerful AI tools in the hands of a few nations or organizations. Several major technology companies, including Nvidia, Microsoft, and Amazon, have urged the White House not to impose similar bans on Chinese open-weight models.

OpenAI's cautious approach with Astra reflects a broader industry challenge: how to make powerful AI tools widely available while preventing misuse. The company's decision to involve government agencies and safety organizations in testing suggests that decisions about Astra's release may ultimately involve regulatory input, not just internal company judgment.

OpenAI has not yet disclosed whether Astra will become part of the GPT-5.6 family of models or mark the beginning of a new GPT-6 generation. According to reports, CEO Sam Altman has already showcased the model to federal officials, indicating that government stakeholders are being kept informed of the model's progress and safety status.