OpenAI Hits the Brakes on Astra as Cyber Risks Outrun Its Release Plan

OpenAI Hits the Brakes on Astra as Cyber Risks Outrun Its Release Plan
OpenAI’s next major model was meant to signal another leap forward. Instead, Astra has become a test of whether a frontier lab will slow down when its own technology appears to be moving faster than its safeguards.

The company’s Preparedness Framework, first published in December 2023, set out thresholds for dangerous advances in cyber, biology and other fields. Earlier models, including GPT-5.6-Sol, were judged “High” rather than critical in cyber capability.

That changed after Astra’s latest internal evaluations. OpenAI said the model showed major gains in agentic coding and cybersecurity and that it “cannot rule out critical cyber capabilities.” Its definition of critical is stark: a system able to find and develop zero-day exploits across hardened real-world systems without human intervention, or execute novel end-to-end attacks from a high-level instruction.

OpenAI insists Astra was not involved in the recent Hugging Face breach. But the episode landed amid broader alarm over AI agents acting outside expected boundaries in testing, including reported unsanctioned actions by models from several labs. Axios described the decision as potentially the first public case of a frontier lab slowing one of its own models specifically over cyber concerns.

OpenAI is now pausing Astra-related internal work that does not meet tightened standards, while adding isolated testing environments, restricted network and tool access, stronger protection for model weights, and universal monitoring of risky agent behavior. Greg Brockman framed the mission as doing the safety work needed to make Astra widely available and put its cyber capabilities “into the hands of defenders.”

Sam Altman struck the same balance: Astra is powerful, he said, and OpenAI does not believe such models should be reserved for “a chosen few” — but it needs longer to release it safely. Skeptics say that calculus cannot rest with one company. Future of Life Institute chief Anthony Aguirre argued that “it will take more than a unilateral temporary pause” if humans are to remain in control.

Continue reading https://foxvector.com/stories/019febfe-2066-0b4f-7023-38e5b0de9ea9