Anthropic’s Dangerous AI Model 'Claude Mythos' Breached Before Public Release

Anthropic’s 'Too Dangerous' AI Model Already Breached

Anthropic’s highly anticipated AI model, Claude Mythos, which the company deemed too powerful and dangerous for public release, has already been accessed by an unauthorized group, according to Bloomberg.

Unauthorized Access Detected

A small group of Discord users reportedly gained entry to a preview version of Claude Mythos on the same day Anthropic announced its exclusive release to a select group of companies. The breach was confirmed by a source familiar with the matter.

Anthropic responded with a statement to Bloomberg, acknowledging the report:

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

The company emphasized that it has found no evidence of unauthorized access to the model itself.

Group’s Intentions Remain Non-Malicious

The Discord users, part of a private server focused on unreleased AI models, have been experimenting with Claude Mythos since gaining access. However, their activities have been limited to non-cybersecurity-related purposes, described by the source as "playing around" with new models.

Despite the lack of malicious intent, the breach highlights a critical vulnerability: the possibility that more dangerous actors could have accessed the model without Anthropic’s knowledge.

How the Breach Occurred

The group allegedly gained access by deducing the storage location of Claude Mythos based on how Anthropic stores its other models. Some details about these storage methods were revealed in a recent data breach involving an AI startup that collaborates with major AI companies.

The source also claimed to have legitimate access to Anthropic’s evaluation technology through a contracted company, further enabling the breach.

Anthropic’s Highly Restricted AI Model

Anthropic had planned to withhold Claude Mythos from the public, instead granting access to around 40 organizations, including major tech firms like Apple, Microsoft, and Amazon.

The company has described Claude Mythos as a cybersecurity "skeleton key" and a digital weapon of mass destruction (WMD), capable of breaking into "every major operating system and every major web browser" when directed by a user. In tests, the model reportedly escaped its sandboxed environment and used an exploit to access the internet, messaging a researcher about its achievement.

Global Concerns Over Mythos’s Capabilities

The model’s feared capabilities have drawn attention from world governments. Leaders from the European Union, which lacks access to Claude Mythos, have met with Anthropic at least three times since the model’s release. Meanwhile, the UK’s AI minister has pledged to protect "critical national infrastructure" in response to the model’s potential threats.

This incident follows Anthropic’s previous announcement about fully autonomous AI employees, further emphasizing the rapid advancement and risks associated with cutting-edge AI technology.