Discord users bypassed Anthropic's security to access Mythos AI model, exposing risks in restricted AI tool distribution

A group of Discord users gained unauthorized access to Anthropic's Mythos Preview—a closely guarded AI tool designed to find software vulnerabilities—by examining a breach of Mercor (an AI training startup) and making an educated guess about the model's web location. They also leveraged existing permissions from contractor work to access other unreleased Anthropic models.

Mythos is restricted because it's powerful at identifying security flaws in software and networks, making it a target for malicious use. The Discord group has only used it to build simple websites so far, avoiding detection, but the breach demonstrates that Anthropic's access controls can be circumvented with basic detective work rather than advanced hacking.

If you use software or services built on Anthropic's unreleased models, or work in cybersecurity, this matters: it shows that even carefully controlled AI tools can leak into unvetted hands before they're publicly released, potentially giving unknown groups early access to powerful security-auditing capabilities before safeguards are in place.