Unauthorized Users Gain Access to Anthropic Mythos AI Model

Unauthorized users have gained access to Anthropic Mythos, an advanced AI model described by Anthropic as highly powerful and potentially risky if misused. The group claims to have accessed the tool since its initial announcement. This incident comes as technology firms seek to secure their systems before releasing such tools more widely.

Key Highlights

• Unauthorized users accessed Anthropic Mythos AI model via a third-party contractor account
• Anthropic is investigating the breach and found no evidence of access beyond the third-party vendor
• The group has not used Mythos for cybersecurity purposes and claims interest in exploring new models
• Project Glasswing provided Mythos to select software firms for security testing before public release

Anthropic has not made Mythos available to the public. Instead, it distributed the model to select software providers through Project Glasswing. This initiative aims to help these firms test and strengthen their defenses against cyberattacks. According to a Bloomberg report, the group accessed Mythos through a third party, using tactics that included leveraging the account of a contractor working for Anthropic.

Details of the Security Breach

Anthropic confirmed it is aware of unauthorized access to the Claude Mythos Preview. The company is investigating the incident and stated it has found no evidence that the group accessed systems beyond the third-party vendor. In a statement to Bloomberg, an Anthropic spokesperson said, “We’re investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments.”

When announcing Mythos, Anthropic highlighted the model’s ability to identify and exploit vulnerabilities in every major operating system and web browser, if directed by a user. The group of unauthorized users reportedly participates in a private Discord channel focused on uncovering information about unreleased AI models. Bloomberg reported that the group has used Mythos regularly since their initial access, but not for cybersecurity-related activities.

Implications for AI Security

To access Mythos, the group deduced the model’s online location by analyzing the format Anthropic used for other models. According to Bloomberg, a group member said their interest lies in exploring new models rather than causing harm. The group has not used Mythos for cybersecurity prompts and also has access to other unreleased Anthropic AI models.

This unauthorized access raises concerns about the ability of AI companies to prevent their most advanced technology from spreading beyond trusted partners. It also prompts questions about whether other unauthorized users may have gained access to Mythos and how the tool might be used.