Unauthorized Access to Mythos Preview: What We Know So Far

Key Points

• Reports emerged in late April 2026 of unauthorized access to Claude Mythos Preview
• Access reportedly occurred through a third-party contractor's environment, not Anthropic's primary systems
• A private Discord community focused on unreleased AI models was involved
• Unauthorized users reportedly leveraged techniques from a separate breach (Mercor)
• Anthropic confirmed the investigation but stated no evidence of primary system compromise
• There are no reports that unauthorized users used the model for exploit hunting

What Happened

According to reporting from Bloomberg and other outlets, an unauthorized group of users gained access to Claude Mythos Preview in mid-to-late April 2026. The access occurred not through a direct breach of Anthropic's systems, but through an environment belonging to a third-party contractor that works with Anthropic.

The unauthorized users were reportedly associated with a private Discord community that specializes in accessing unreleased AI models. The group used a combination of common research techniques and intelligence from a separate, unrelated breach involving the AI startup Mercor to identify the model's online location in the contractor's environment.

How the Access Occurred

Based on available reporting, the sequence appears to have been:

1. Information gathering: The group used publicly available information and insights from the Mercor breach to identify potential access points
2. Contractor vulnerability: The third-party contractor's environment had insufficient access controls, creating an entry point
3. Model access: Once inside the contractor environment, the group was able to interact with Claude Mythos Preview
4. Discovery: The access was identified and reported to Anthropic

The reporting emphasizes that this was not a sophisticated cyberattack against Anthropic's core infrastructure. Rather, it was an exploitation of weak security in a partner's environment — a common vector for security incidents across the technology industry.

Anthropic's Response

Anthropic has confirmed that it is investigating the reports. In its official statement, the company noted:

• There is no evidence that the unauthorized activity extended beyond the third-party vendor's environment
• Anthropic's own primary systems were not compromised
• The company is conducting a thorough review of third-party access controls
• Mitigation measures have been implemented

What the Unauthorized Users Did (and Didn't Do)

One important detail from the reporting: there are no indications that the unauthorized users attempted to use Mythos Preview for its intended cybersecurity capabilities. Reports suggest they did not use the model to hunt for new exploits or vulnerabilities.

This is notable because the primary concern around unauthorized Mythos Preview access has always been the potential for the model's exploit-discovery capabilities to be used maliciously. The apparent absence of such use, while not exonerating the access, provides some reassurance.

Implications for Project Glasswing

This incident raises questions about the security of the Project Glasswing access model:

• Third-party risk: Even tightly controlled access programs are only as secure as their weakest participant
• Contractor oversight: The incident highlights the need for rigorous security requirements for all entities in the Mythos Preview access chain
• Access monitoring: Whether Anthropic had adequate real-time monitoring to detect unauthorized use remains an open question

Status

As of April 24, 2026, the investigation is ongoing. We will update this article as new information becomes available. For context on the model involved, see our explainer and system card analysis.