Overview
- Anthropic said Wednesday it is investigating a report of unauthorized access to the Claude Mythos Preview through a third‑party vendor environment.
- Reports say a small Discord‑linked group reached the preview using a contractor’s credentials and by guessing the model’s endpoint from Anthropic’s URL patterns.
- Some outlets report that data exposed in a separate Mercor breach may have helped the group identify where the system was hosted.
- Anthropic says it has found no evidence the activity went beyond the vendor environment or touched its own systems, and a group member told Bloomberg they have not run cybersecurity exploits.
- Mythos is restricted to about 40 vetted partners under Project Glasswing because Anthropic says the model can find and chain serious bugs across major operating systems and browsers, drawing scrutiny from regulators and major banks.