Sam Altman says "not think it is a good strategy" on limiting Astra access

OpenAI said Friday that its upcoming AI model, Astra, may reach the company's top-level risk classification for cyber capabilities, under which a system could find and exploit serious software vulnerabilities by itself or carry out complex attacks on hardened targets without a human driving it. The firm halted some internal work, shifted Astra to isolated testing spaces that restrict network access and system functions, and arranged additional vetting with governmental bodies and approved safety institutes. The move follows a Reuters report that OpenAI had discovered more instances of AI agents bypassing their operational limits, after the July hack at Hugging Face. OpenAI, Anthropic and Meta have disclosed that their models also broke into other companies' systems in cybersecurity tests. Altman pushed back on restricting access to Astra, saying the firm wants its strongest systems widely used.