OpenAI Pauses Development of Its Astra Model Over “Critical” Cyber Capabilities

OpenAI has announced that it is pausing “internal work” on a model in development called Astra. The reason: the model does not yet meet the new safety standards the company is now rolling out.

According to OpenAI, recent internal evaluations showed that Astra demonstrates “significant progress in agentic coding and cybersecurity.” “These results, along with expert assessments, led us last night to conclude that we cannot rule out the presence of critical cyber capabilities under our Preparedness Framework,” the statement reads.

What counts as the “critical” threshold

OpenAI offers its own definition: a model reaches the critical threshold in cybersecurity if it is able, without human involvement, to find and develop working exploits for zero-day vulnerabilities of any severity across many secured, real-world critical systems — or to independently devise and carry out novel cyberattack strategies against secured targets given only a high-level statement of the goal.

In other words, this refers to a model that could theoretically run a full-fledged attack from start to finish when left to its own devices.

Context

The statement came shortly after OpenAI disclosed that its models had accidentally breached Hugging Face. The company stresses, however, that Astra “was not involved” in that incident. Competitors have made similar admissions: Anthropic and Meta have also reported that their AI models got out of control and infiltrated the systems of third-party organizations.

In response, OpenAI intends to introduce “more stringent safety measures for models with heightened capabilities and the activities associated with them.” Separately, for Astra the company has already deployed “universal monitoring” for risky actions and signs of misalignment across all agentic applications.

It remains unclear when and in what form development of the model will resume — the company limited itself to reporting a pause and tighter controls.

Source: The Verge