Anthropic Prepares Mythos Model for Claude Code Rollout

Understanding Anthropic’s Powerful, Risky Mythos Model

The Mythos model represents a significant leap in AI’s ability to interact with and analyze code. Its core advancement lies in its striking proficiency in computer security tasks, including the automatic development of functional cyberattacks at a professional level. Anthropic itself has cautioned that this model poses a severe risk to global digital infrastructure if not handled with extreme care.

The company explicitly warned about the potential for attackers to exploit these tools in the short term if frontier labs are not diligent in their release strategies. Conversely, Anthropic anticipates that in the long term, defenders will leverage such models to more efficiently identify and fix bugs before new code is deployed. This dual-edged capability underscores the critical balance Anthropic is attempting to strike between innovation and security.

Anthropic’s Strategic Approach to Mythos Deployment

To mitigate the significant security risks associated with Mythos, particularly the potential for exploiting unpatched vulnerabilities in popular applications, Anthropic has adopted a phased and cautious deployment strategy. The company has prioritized the development of robust guardrail systems before any public rollout.

1. Anthropic initially announced the Mythos model in early preview on April 7, calling it a new frontier model with advanced capabilities in computer security tasks.
2. The company recognized the model’s potential to automatically develop functional cyberattacks, leading to a decision against immediate public release due to severe risks to global digital infrastructure.
3. Anthropic focused on developing a powerful guardrail system to prevent malicious exploitation of the model’s capabilities.
4. References to the model, specifically claude-mythos-1-preview, have since appeared within Anthropic’s Claude Code and Claude Security environments, indicating internal integration and testing.
5. Briefly, some users observed a toggle to enable Mythos within the public version of Claude Code before it was subsequently taken offline, suggesting ongoing internal preparations for a broader launch.
6. Anthropic’s current offerings include Claude Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, and Haiku 4.5, with Mythos slated to join this lineup once deemed secure for wider release.

Project Glasswing: Securing Digital Infrastructure with Mythos

While the full public rollout of Mythos is pending, Anthropic has already begun leveraging its capabilities in a controlled environment through “Project Glasswing.” This initiative involves collaborating with other organizations to secure critical software against potential AI-driven exploits. Anthropic has confirmed its work on this project, which utilizes the unreleased Claude Mythos Preview.

Project Glasswing has already assisted approximately 50 organizational partners, demonstrating the model’s defensive potential. Anthropic showcased a dashboard revealing a multitude of open-source vulnerabilities identified by Mythos Preview. In its first month alone, the Mythos model managed to uncover 10,000 high- or critical-severity vulnerabilities, underscoring its efficacy as a defensive tool. This impressive capability helps explain Anthropic’s cautious approach to public release, prioritizing the development of robust safeguards.