Discord Users Gained Unauthorized Access to Anthropic’s Restricted Mythos AI Model

A group of amateur investigators on Discord obtained unauthorized access to Anthropic’s Mythos Preview, an AI model the company has tightly restricted because of its ability to find security vulnerabilities. The users gained entry through straightforward detective work rather than advanced hacking, according to Bloomberg. They examined data from a recent breach of Mercor, an AI training startup, and made an educated guess about the model’s online location based on the format Anthropic uses for other models. One person also leveraged existing permissions from work at an Anthropic contracting firm. The probing allegedly exposed not only Mythos but also other unreleased Anthropic models. Bloomberg reported that the group has so far used the access only to build simple websites, a choice intended to avoid detection by Anthropic. Mozilla said on Tuesday that it used legitimate early access to Mythos Preview to find and fix 271 vulnerabilities in its new Firefox 150 browser release. Anthropic has marketed Mythos as a dangerously capable tool for uncovering flaws in software and networks.