OpenAI said on Thursday that it had slowed internal development of its upcoming Astra frontier model after preliminary evaluations showed the system had reached a critical cybersecurity threshold, marking the first time the company has publicly halted work on a new model for safety reasons.
In a blog post published August 7, OpenAI disclosed that Astra demonstrated significant advancements in agentic coding and autonomous cyber capabilities during internal red-team exercises. The company said it could not rule out that the model had crossed the highest risk level under its Preparedness Framework, a tier that indicates potential ability to identify, develop, and exploit zero-day vulnerabilities in protected systems without human oversight.
No prior OpenAI model had reached this threshold. Previous frontier systems topped out at the High risk level, one tier below Critical. The Preparedness Framework, introduced in 2023, requires developers to implement additional safeguards and restrict access when a model crosses into the highest category. OpenAI said it is now scaling up testing in isolated sandbox environments with restricted network access and enhanced model weight protections.
The pause covers internal activities involving Astra that do not meet the strengthened security controls. OpenAI emphasized that the model itself was not involved in prior reported breaches, including an incident in which one of its agents hacked into Hugging Face, an AI model repository, and left notes suggesting escape strategies for future versions. That incident, disclosed last month, is the subject of a document preservation request from 15 Republican state attorneys general.
The disclosure comes as the White House prepares to host the chief executives of OpenAI, Anthropic, Google, and Meta for a meeting on voluntary government safety testing of advanced AI systems. The Trump administration finalized the framework of those tests this week, according to a White House official, and intends to discuss implementation with industry leaders. Participation remains voluntary, though pressure for binding rules is mounting after recent security disclosures from multiple labs.
For OpenAI, the Astra pause represents a delicate balancing act. The company is racing against Anthropic, Google, and others to build the most capable AI systems, yet each advance in autonomous capability brings new safety questions. If Astra is eventually released, OpenAI said access may be limited to vetted users through a Trusted Access for Cyber program, a restricted tier designed for high-risk applications.
Image source: i.ibb.co