Anthropic is investigating a reported case of unauthorized access to its Claude Mythos model, an AI tool designed to identify cybersecurity vulnerabilities, the company confirmed to Bloomberg.

A group reportedly gained entry to the model through a third-party contractor portal and by using internet sleuthing tools, according to sources cited by the report. However, a person familiar with the matter stated that the group’s primary interest was testing the model rather than using it for malicious purposes.

In an official statement, Anthropic said:

"We're investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments."

The Claude Mythos Preview was released earlier this month as part of Project Glasswing, initially limited to a select group of trusted test companies. These included Amazon, Microsoft, Apple, Cisco, and Mozilla.

Mozilla disclosed that the model helped identify and patch 271 Firefox vulnerabilities. The tool’s capabilities have also drawn interest from banks and government agencies seeking to enhance their own security measures.

Despite its restricted access, several unauthorized users—allegedly part of a private Discord chat—claimed to have accessed Mythos through a developer portal. They reportedly deduced the model’s location through educated guesswork. Sources suggest the group may have also gained access to other unreleased Anthropic models.

The Mythos model has recently gained attention for its supposed ability to detect security flaws in operating systems and web browsers. This has sparked both skepticism among security researchers and concerns about the potential for AI-driven cyberattacks becoming a significant threat, according to Alex Zenla, CTO of cloud security firm Edera, who spoke to Wired.

Anthropic was recently designated as a "supply chain risk" by the U.S. Department of Defense, though the company has been in discussions with the Trump administration to have this label reconsidered.

Source: Engadget