OpenAI has paused certain internal work on its upcoming Astra model after preliminary evaluations found the system’s agentic coding and cybersecurity abilities advanced enough that a “Critical” capability classification under the company’s Preparedness Framework could not be ruled out. Under that framework, “Critical” means a model could independently discover previously unknown software vulnerabilities or plan and execute sophisticated cyberattacks with minimal human guidance. OpenAI says it has since applied stricter security controls, including isolated testing environments, restricted network and tool access, and additional monitoring, and paused any Astra-related activity that does not meet those enhanced requirements. Before deployment, the company plans to have government agencies and independent safety organizations separately evaluate Astra’s cybersecurity capabilities.
