OpenAI slows Astra model development over cybersecurity concerns

OpenAI said Wednesday it has slowed development of its Astra model after the system, still in training, crossed what the company calls its “critical cybersecurity threshold.” According to OpenAI, that designation means Astra demonstrated the capacity to independently identify and carry out cyberattacks against real-world systems that are traditionally well-protected, TechCrunch reported August 7.

The company did not disclose the specific capabilities that triggered the threshold or provide technical detail on how Astra’s performance was evaluated. OpenAI has not given a timeline for resuming full-speed development, characterizing the pause as a response to internal safety assessments rather than an external mandate.

Astra remains unreleased. TechCrunch’s report does not indicate whether any version of the model has been deployed in limited testing or made available to outside researchers.

The disclosure marks one of the first instances of a major AI developer publicly attributing a slowdown in a flagship model’s development to a specific, named safety threshold tied to offensive cyber capability, rather than to general risk language. OpenAI has previously described threshold-based safety frameworks for its models in broad terms; this report ties that framework to a concrete case for the first time.

No further detail on next steps, review timelines, or the mechanics of the threshold has been published as of this writing.