OpenAI said Friday it is suspending some internal research and development work on its next-generation frontier AI model, Astra, after internal evaluations found the system may have "critical-level cyberattack capabilities" and that parts of the work had not met enhanced security controls. The company said Astra showed significantly stronger coding and cybersecurity performance, and its assessments could not rule out autonomously identifying and exploiting zero-day vulnerabilities without human involvement. A newer report described those capabilities as potentially warranting a "dangerous" rating. OpenAI said the move is a pause on activities that have not met tighter safeguards, not a full halt, while it upgrades development and testing controls and works with government agencies and AI safety groups on boundary testing. Separately, OpenAI opened unlimited text-only access to its faster but lower-performance GPT-5.6 Luna model for ChatGPT free users and users on its lower-priced Go plan, while improving factual reliability in GPT-5.6 Sol for Plus and Pro users. The Astra pause follows recent containment failures across AI labs, including July incidents involving OpenAI, Anthropic and Meta, and adds to expectations of higher spending on testing, monitoring, insurance and compliance as the cybersecurity industry shifts toward AI-native security.