## Anthropic's Claude Mythos leaked to unauthorized users despite safety claims
Anthropic's controlled rollout of Claude Mythos has suffered a significant setback. The AI company spent weeks claiming the model posed such serious cybersecurity risks that it could not be released publicly—yet a small group of unauthorized users appears to have accessed the system since the day Anthropic announced plans to offer it to a select group of enterprise partners for testing. Anthropic has acknowledged the situation and says it is investigating how the breach occurred.

The existence of Claude Mythos was itself first revealed through an external leak, before Anthropic officially announced the model as part of its cybersecurity-focused offerings. The company positioned the model's advanced capabilities as a double-edged sword: too potent for public release but valuable enough to justify restricted enterprise access. That narrative has now been complicated by reports that the very controls Anthropic promised were already compromised at launch. The incident raises questions about the company's internal security protocols and its ability to manage access to models it deems dangerous enough to restrict.

The breach carries reputational weight beyond the immediate technical exposure. Anthropic has built its brand on rigorous AI safety commitments and careful deployment practices, positioning itself as a more cautious alternative to competitors. A leak of this nature—particularly one involving a model explicitly described as a security risk—undermines that positioning and invites scrutiny from regulators, partners, and safety researchers already skeptical of frontier AI companies' self-governance claims. Anthropic has not disclosed how many users accessed the model or what they were able to do with it.
---
- **Source**: The Verge
- **Sector**: The Lab
- **Tags**: anthropic, claude mythos, ai leak, ai safety, cybersecurity
- **Credibility**: unverified
- **Published**: 2026-04-23 23:24:07
- **ID**: 76548
- **URL**: https://whisperx.ai/en/intel/76548