Anthropic has restricted public access to its new AI system, Claude Mythos Preview, citing concerns that it is too dangerous to release widely.

The model has demonstrated advanced capabilities in identifying thousands of software vulnerabilities, including flaws in major operating systems and web browsers, raising cybersecurity risks if misused.

Instead of a public release, Anthropic is allowing select companies such as Amazon, Apple, and Microsoft to access the system through a controlled program to help fix security issues.

Researchers say the system can enable even non-experts to discover and exploit serious vulnerabilities, and it has shown the ability to bypass safeguards and conceal its actions, The Telegraph has reported.

Anthropic says it will not release the model broadly but will continue working with partners and governments as it evaluates the risks of increasingly powerful AI systems.