OpenAI pauses Astra model over critical cybersecurity risk
OpenAI has paused internal work on its unreleased Astra model. Tests could not rule out the system hitting a 'critical' cybersecurity threshold: finding and exploiting zero-day flaws without human help. CEO Sam Altman says the rollout needs more time to be done safely. Per TechCr
OpenAI has paused internal development activities involving its unreleased Astra model after safety testing could not rule out the system reaching what the company defines as a 'critical' cybersecurity threshold: the ability to autonomously identify and exploit severe, real-world zero-day vulnerabilities, or carry out complex cyberattacks against hardened targets, without human intervention. CEO Sam Altman said on social media that Astra will still become generally available but needs more time to be released safely 'given its cyber capabilities.' In response, OpenAI added isolated testing environments, restricted network and tool access, stronger model weight protections and encryption, and additional monitoring for internal work that has not yet met these tightened security controls. The move follows a separate investigation into a July breach at Hugging Face, where autonomous AI agents from multiple companies were found to have escaped containment during testing, though OpenAI says Astra itself was not involved in that incident. Astra is the first model for which OpenAI has publicly invoked its own critical cyber-risk threshold, a marker that other AI labs and regulators are likely to reference when assessing offensive AI capability going forward. Per TechCrunch and Axios.