Anthropic Limits Access to New AI Model on Cyberattack Misuse Concerns

Anthropic has restricted access to its latest AI model, citing security risks.

Cointelegraph reported on April 7 that the company would release its new model, Claude Mythos Preview, only to a limited number of companies.

The model has shown the ability to find thousands of critical vulnerabilities in major software, including operating systems and web browsers. Anthropic said the capability could soon become widely available given the pace of AI development, raising concerns about misuse.

AI-driven cyberattacks are already increasing rapidly. Data cited in the report showed 87% of organizations worldwide experienced AI-enabled attacks in 2025, while the number of attacks rose 72% from a year earlier.

Anthropic has launched Project Glasswing to address those risks. More than 40 companies, including Amazon Web Services, Apple, Google and Microsoft, plan to take part in building AI-based systems for vulnerability detection and response.

Claude Mythos Preview also identified vulnerabilities that had gone undetected for more than a decade. It found a 27-year-old flaw in the OpenBSD operating system and uncovered multiple security defects in FFmpeg, FreeBSD and the Linux kernel.

The model also found weaknesses in major encryption technologies, including TLS, AES-GCM and SSH. Anthropic said 99% of the vulnerabilities it identified remain unpatched and withheld detailed information.

Anthropic said AI will improve security over the long term, but threats could expand during the transition period.