Anthropic plans wider rollout for Mythos

Saturday, June 6, 2026

Anthropic is expanding access to its AI tool, Claude Mythos Preview, which helps detect software vulnerabilities, increasing its Project Glasswing partnership from 50 to around 200 organizations. This expansion is significant as it supports critical sectors like government and finance in addressing cybersecurity risks highlighted by the discovery of over 10,000 severe security flaws.

Anthropic announced on Tuesday plans to expand access to its artificial intelligence tool aimed at identifying software vulnerabilities, a move that has drawn significant attention from government agencies and financial institutions globally. The San Francisco-based company revealed that its Project Glasswing, initially a partnership involving about 50 organizations, will grow to encompass approximately 200 partners.

Project Glasswing grants qualified members access to Anthropic’s AI system, known as Claude Mythos Preview. Each organization must meet specific security criteria to participate, according to Anthropic. The company worked closely with the U.S. government, current Glasswing partners, and other stakeholders in the cybersecurity sector to facilitate the broader rollout.

The announcement follows the company’s recent filing for what could become one of the largest initial public offerings (IPOs) to date, underscoring heightened commercial and governmental interest in its AI capabilities. Anthropic’s Mythos Preview, unveiled in April, demonstrated an advanced ability to detect ways that seemingly minor software bugs could be exploited in combination, raising alarms about potential hacking risks.

This revelation prompted a swift response from top U.S. officials, including Treasury Secretary Scott Bessent and then-Federal Reserve Chair Jerome H. Powell, who held an emergency meeting with bank leaders to discuss the emerging threats. Financial institutions have since sought access to the AI tool to identify and address vulnerabilities swiftly.

Anthropic reported that its partners have uncovered over 10,000 security flaws deemed highly or critically severe since the AI’s deployment. However, some experts in the cybersecurity field caution that concerns about widespread, uncontrolled hacking enabled by the technology may be exaggerated.

The company’s expanding partnership and cautious approach to access reflect ongoing efforts to balance the potential benefits of advanced AI-driven security analysis with the risks inherent to powerful new tools in digital defense.