A select group of unauthorized individuals has managed to gain entry into Anthropic Mythos, an advanced AI model considered potent enough to pose a significant hacking risk if misused. These users claim to have had access to the tool since its initial announcement. This development coincides with a time when numerous companies are keen on fortifying their systems against potential threats before making such tools widely available.
Although the tool is not accessible to the general public, Anthropic has made it available to a limited group of software providers through Project Glasswing. This initiative was designed to enable these companies to test and secure their systems against potential cyber threats.
According to a report by Bloomberg, the group obtained access to the tool through a third party, utilizing various tactics, including leveraging the credentials of an individual associated with a third-party contractor for Anthropic.
Anthropic has acknowledged the unauthorized access to Claude Mythos Preview and is currently conducting an investigation into the matter. The company has stated that there is no evidence indicating that the group has breached any systems beyond those of the third-party vendor or Anthropic.
In response to the situation, a spokesperson for Anthropic informed Bloomberg that they are looking into the reported unauthorized access to Claude Mythos Preview via a third-party vendor environment.
The capabilities of the Anthropic Mythos tool have raised concerns, as it was initially presented as having the ability to identify and exploit vulnerabilities in major operating systems and web browsers upon user direction.
The users involved are part of a private Discord channel dedicated to discovering information about unreleased models. They have reportedly been actively using Mythos since gaining access, albeit not for cybersecurity purposes.
To access Mythos, the user group made an educated guess about the model’s online location based on their understanding of the format typically used by Anthropic for other models. The group has emphasized their interest in exploring new models rather than causing harm, highlighting that they have not engaged in cybersecurity-related activities with the Mythos model. Additionally, they have access to several other unreleased Anthropic AI models.
The unauthorized access to Mythos has prompted discussions about the challenges faced by AI companies in safeguarding their most powerful and potentially risky technologies from being misused by unauthorized parties. This incident also raises questions about whether other individuals may be using Mythos without authorization and for what purposes.