Anthropic Limits Mythos AI Rollout Over Cyberattack Risks

Anthropic announced on Tuesday that it is developing a new advanced AI model called Mythos , which will be shared with a limited group of companies as part of a cybersecurity initiative called Project Glasswing. The idea behind this project is to explore how powerful AI can be used to strengthen digital security without creating unnecessary risks.

The model, called Claude Mythos Preview, is particularly strong at identifying weaknesses and security flaws in software systems. It can scan complex code and spot issues that might otherwise go unnoticed. Because of how capable the model is, Anthropic is being very cautious about who gets access to it. The company wants to avoid a situation where such a powerful tool could be misused by bad actors.

Several major tech companies are part of the initial rollout. This includes Apple, Google, Microsoft, Nvidia, and Amazon Web Services. These companies will use the model mainly for defensive security work, such as identifying vulnerabilities and improving system safety. In addition to these tech giants, more than 40 other organizations, including cybersecurity firms like CrowdStrike and Palo Alto Networks, are also participating in the program.

An interesting thing about Claude Mythos Preview is its ability to uncover bugs that have been hidden for 27 years, one example being an operating system called OpenBSD which has always been very focused on security. This shows how AI can find errors that developers may not be able to find after years of working.

Anthropic indicated the model was not exclusively trained for cyber-security. Rather, the model’s high level performance for cyber-security was due to its capabilities in reasoning and programming. The model is considered a general software development tool that can be applied to a variety of technology-related issues. While the company views the model as capable of performing many functions, it does not anticipate releasing it to the public in the near future. Currently, the focus is on developing a strategy for how to scale and safely implement this type of advanced system into everyday use.

The companies involved in Project Glasswing are responsible for building or maintaining critical software infrastructure. They will use the model to improve both their own systems and open-source projects. To support this effort, Anthropic has committed up to $100 million in usage credits. Once companies cross that limit, they will need to pay for continued use.

Newton Cheng, who leads Anthropic’s cybersecurity testing team, explained that the goal is to help organizations get comfortable with these advanced tools before they become widely available. He emphasized that the company wants to avoid releasing something too powerful without proper safeguards in place.