Project Glasswing
Also known as: Glasswing, Anthropic Glasswing
Project Glasswing launched alongside the Claude Mythos Preview announcement in April 2026. Rather than releasing a model it considered too capable for unrestricted use, Anthropic structured a narrow deployment: a vetted group of critical infrastructure companies, cloud providers, and security organizations got early access under terms that restrict usage to defensive cybersecurity. In return, they share findings with Anthropic as safety telemetry.
The practical work inside Glasswing involves using Mythos to scan codebases for zero-day vulnerabilities (software flaws unknown to developers), triage severity, and coordinate responsible disclosure to open-source maintainers and closed-source vendors. Anthropic reported finding thousands of such vulnerabilities in major operating systems and browsers in just a few weeks of testing. Named participants have included AWS, Apple, Cisco, Google, JPMorgan Chase, and Microsoft.
For builders without Glasswing access, the term matters for two reasons. First, it introduced a new governance pattern: a gated safety consortium as an alternative to full public release or full withholding, which other labs are likely to adopt. Second, it signals that AI-powered security scanning is becoming a first-class capability, meaning both offensive and defensive postures around AI-generated code are about to change significantly.