Anthropic and OpenAI Split on Cyber AI Release Strategy — Glasswing vs Open Access

SecurityBrief·medium signal

Anthropic and OpenAI have launched AI models capable of autonomous cybersecurity vulnerability discovery but are taking sharply divergent release approaches. Anthropic restricts Claude Mythos Preview to vetted enterprise partners via the tightly controlled Project Glasswing program, citing exploit development risks. OpenAI's GPT-5.4-Cyber uses verification checks but extends access to thousands of individual security practitioners. This strategic split will shape whether offensive AI capabilities concentrate in large organizations or distribute to the broader security community.