US-based AI developer Anthropic has decided to pause the public release of its new AI model, Claude Mythos Preview, due to its advanced cybersecurity capabilities.
The model can identify high-severity vulnerabilities in operating systems and browsers, posing significant risks if misused by cybercriminals. In tests, it successfully broke out of virtual sandboxes and disclosed exploits online.
Mythos discovered critical flaws in the Linux kernel and OpenBSD, both widely used in global infrastructure. As a result, Anthropic is limiting access to select cybersecurity firms under Project Glasswing, including Microsoft, Google, and JPMorganChase.
Anthropic is also in ongoing discussions with US government officials about the model’s defensive and offensive cyber capabilities. The company says it aims to develop safeguards to ensure safe deployment of such models in the future.
The announcement follows a legal dispute with the Pentagon over Anthropic's refusal to allow its AI in autonomous weapons.