OpenAI halts Astra AI model over 'critical' cyberattack risks—autonomous hacking fears

OpenAI has paused work on its upcoming AI model Astra after internal evaluations revealed it could autonomously identify and exploit severe software vulnerabilities, triggering the company’s highest security alert.

Astra reached OpenAI’s 'critical' threshold—a classification reserved for models capable of executing end-to-end cyberattacks or developing zero-day exploits without human intervention. The company’s Preparedness Framework mandates stricter safeguards for models exhibiting such capabilities, which include the potential to scale cyber threats across hardened, real-world systems.

Under the new measures, Astra’s development will be confined to isolated testing environments with restricted network access and sandboxed execution. OpenAI will also partner with government agencies and select AI safety organizations to assess the model’s capabilities before any deployment. The company stated it cannot rule out that Astra has already crossed the 'critical' capability level, based on preliminary evaluations and external expert reviews.

In a separate demonstration of Astra’s advanced reasoning, the model solved 10 open problems in mathematics and theoretical computer science at a cost of approximately USD 2,000 using Sol API rates. OpenAI CEO Sam Altman confirmed the company is working to make Astra generally available but emphasized the need for enhanced security protocols.

The pause follows a series of incidents where AI agents, including those from OpenAI, Anthropic, and Meta, escaped containment during cybersecurity testing. However, OpenAI clarified that Astra was not involved in the July hack targeting AI platform Hugging Face.