Explained: Project Glasswing and how it is linked to Mythos Preview’s limited rollout

Anthropic said the project is necessary because modern AI systems have reached a stage where they can detect and exploit software vulnerabilities at a level comparable to, or exceeding, skilled human experts. The company has committed up to $100 million in usage credits and additional funding for open-source security efforts.

Participating organisations are using the model to scan codebases, conduct penetration testing, and strengthen security across systems that form a large portion of the global digital ecosystem.

What is Claude Mythos Preview

Claude Mythos Preview is an unreleased frontier AI model developed by Anthropic. It is described as a general-purpose system with strong coding and reasoning capabilities, particularly in cybersecurity.

During internal testing, the model identified thousands of previously unknown vulnerabilities, including issues in major operating systems and widely used software libraries. In some cases, these vulnerabilities had existed for years without detection.

The model has also demonstrated the ability to autonomously chain together multiple weaknesses to create working exploits, raising concerns about how such capabilities could be misused if widely available.

Why the rollout is limited

The limited rollout of Claude Mythos Preview is directly linked to the risks identified during internal testing. Anthropic has acknowledged that while the model can improve defensive cybersecurity workflows, it also has the potential to enable faster and more effective cyberattacks if misused.

Speaking about the model’s capabilities, Anthropic CEO Dario Amodei said the system was trained primarily for coding but has shown strong cybersecurity performance as a side effect. “We haven’t trained it specifically to be good at cyber. We trained it to be good at code. But as a side effect of being good at code, it’s also good at cyber,” he said, adding that the model is “by and large as good as a professional human at identifying bugs.”

Researchers involved in testing said the model can chain multiple vulnerabilities together to create complex exploits. “It has the ability to chain together vulnerabilities… sometimes three, four, sometimes five vulnerabilities that in sequence give you some kind of very sophisticated end outcome,” one Anthropic researcher explained. This capability, combined with its ability to pursue long-range tasks autonomously, increases the risk profile of the system.

Anthropic has also pointed to the shrinking gap between vulnerability discovery and exploitation. With AI systems accelerating both processes, tasks that previously took months could now be completed in significantly less time, raising concerns about the scale and speed of potential cyber threats.

The company acknowledged these risks directly, stating that “capabilities in a model like this could do harm if in the wrong hands,” which is why it has decided not to release the model widely at this stage.

As a result, access to Mythos Preview has been restricted to selected partners under controlled conditions. These partners are expected to use the model for defensive purposes, including identifying vulnerabilities in critical systems, while also contributing to the development of safety measures and best practices for future deployments.

How Project Glasswing connects to the rollout

Project Glasswing serves as the framework for this controlled deployment. Instead of releasing Mythos Preview publicly, Anthropic is using the initiative to test its real-world impact in a secure and collaborative setting.

The project allows organisations to benefit from the model’s capabilities while ensuring that vulnerabilities are disclosed responsibly and patched before details are made public. It also enables the development of safeguards that can detect and block harmful uses of such AI systems.

Anthropic plans to share findings from the project, including insights into vulnerability detection, patching processes, and evolving cybersecurity practices.

What it means for the future

The link between Project Glasswing and the limited rollout of Mythos Preview highlights a broader shift in AI deployment strategies. As models become more capable, companies are moving toward restricted, use-case-specific releases rather than open access.

For cybersecurity, this could mark the beginning of an AI-driven defence model where vulnerabilities are identified and fixed at scale. At the same time, it underscores the need for strong safeguards to prevent these capabilities from being used for attacks.

Project Glasswing is positioned as an early step in shaping how powerful AI systems are introduced into critical sectors while managing the risks that come with them.