Anthropic expanded vetted organizations’ access to Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 for critical-infrastructure cybersecurity testing.

Anthropic’s reported expansion is restricted to vetted organizations, not unrestricted public access. The available messages identify Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1, with access also intended to encompass future models.

The permitted work includes testing defenses for critical infrastructure. The reporting specifically names power grids, banks, and aviation operating systems, and describes permission for “high-risk offensive testing” (高风险攻击性测试) of their security protections.

Eligibility extends to existing Project Glasswing (玻璃翼计划) participants. These include U.S. government agencies, financial institutions, and major software suppliers, while new organizations are to undergo joint review by Anthropic and the U.S. government.

Additional offensive-testing permissions remain subject to restrictions. The messages say conduct capable of causing physical harm or large-scale destruction remains prohibited, while certified defensive cybersecurity teams may undertake malware reverse engineering and incident response; a separate message describes three access tiers, including red-team and specially vetted access.