Anthropic is working with the U.S. government to broaden access to its most advanced AI models, allowing a screened group of institutions to test the startup's cutting-edge cyber capabilities.

According to a statement released on Tuesday, the San Francisco-based company will permit verified organizations to use its most capable models, including Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1, as well as new models introduced in the future.

The statement said authorized institutions will be able to conduct "high-risk offensive testing" on security systems designed to protect critical infrastructure such as power grids, banks, and flight operations systems.

In April of this year, Anthropic limited access to the Mythos model for U.S. government agencies, financial institutions, and large software vendors through "Project Glasswing."

The release of Mythos became a watershed moment in the cybersecurity field, as the model is capable of automating part of the work of advanced hackers.

Since then, a series of incidents in which models from Anthropic and OpenAI inadvertently intruded into government, corporate, and nonprofit systems have raised concerns about the safety of this technology.

Anthropic said on Tuesday that the latest round of limited access will cover all members of Project Glasswing, and that the company will work with the U.S. government to vet all newly joining institutions.