Descriptions:
Anthropic has officially announced Project Glasswing and Claude Mythos Preview, a new model the company describes as a significant jump in autonomous cybersecurity capability. Published directly on Anthropic’s channel, the video explains that Claude Mythos Preview was not purpose-built for security research—its capabilities emerged as a byproduct of deep code understanding—but it now performs at or near professional human levels for identifying, chaining, and exploiting software vulnerabilities.
In early testing conducted under Project Glasswing, Anthropic’s researchers found a 27-year-old crash vulnerability in OpenBSD (triggerable remotely with minimal data), multiple Linux privilege-escalation bugs allowing unprivileged users to gain root access, and vulnerabilities across other major platforms. All findings were responsibly disclosed to maintainers, who have since shipped patches. The model’s ability to autonomously chain three to five separate vulnerabilities into a single sophisticated exploit is highlighted as a particularly novel capability compared to prior tools.
Because of the dual-use risks, Anthropic is not releasing Claude Mythos Preview publicly. Instead, Project Glasswing gives select partner organizations—those maintaining critical infrastructure and widely used open-source software—early access to use the model defensively. Anthropic has also briefed US government officials and offered collaboration on risk assessment. The initiative positions proactive AI-assisted vulnerability discovery as a necessary response to a world in which increasingly capable models will be available to both defenders and adversaries.
📺 Source: Anthropic · Published April 07, 2026
🏷️ Format: Keynote Launch
