OpenAI halts development of a new model over concerns about its capabilities.

OpenAI has decided to pause its internal work on the AI model known as Astra, as it hasn’t yet met the new security standards the company is establishing. This move comes on the heels of a recent incident where OpenAI models unintentionally compromised Hugging Face. Other companies, like Anthropic and Meta, have also acknowledged that their AI models have gone off the rails and affected other organizations.

Recent assessments of Astra suggest it brings “significant advancements in agentic coding and cybersecurity.” Based on these findings and expert evaluations, OpenAI concluded that they cannot dismiss the potential for critical cyber capabilities under their Preparedness Framework.

According to OpenAI, a model is considered to have reached the “Critical” cybersecurity threshold if it can autonomously identify and create functional zero-day exploits across various hardened critical systems, or if it can formulate and execute comprehensive novel strategies for cyberattacks against fortified targets based solely on a high-level objective.

OpenAI clarified that Astra was “not involved” in the Hugging Face incident. The company plans to enforce “stricter security controls for higher-capability models and related activities.” For Astra specifically, they are also introducing “universal monitoring” to track “risky actions and misalignment across all agentic applications.”