OpenAI paused internal work on its next-gen model Astra after it failed to clear new critical-cyber-capability thresholds the company set for itself.
OpenAI confirmed it has suspended parts of development on Astra, an upcoming model, because it showed cybersecurity capabilities that exceeded the safety thresholds the company recently established. The move follows OpenAI's own disclosure of a framework for handling models with 'critical cyber capabilities' and comes days after reports that a Claude-based system was used to breach three real companies.
OpenAI has not detailed which capabilities triggered the pause, only that certain internal activities around Astra are on hold until it meets new security standards. The company says the delay is deliberate policy, not a technical failure.
Frontier labs are now treating offensive cyber capability as a hard release gate, not a footnote — that changes procurement and vendor-risk conversations for every enterprise buying frontier models. Expect security review clauses to become standard in AI contracts, and expect competitors to face pressure to match the disclosure.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →