Anthropic’s most powerful AI model reportedly accessed in security breach

A small, unauthorised group has gained access to Anthropic’s Mythos AI model, a robust cybersecurity tool said to be potentially harmful if misused, as reported by Bloomberg. 

The group, comprising of a member described only as “an Anthropic contractor,” reportedly tapped into Mythos through a blend of methods, leveraging the contractor's clearance and “standard internet detective techniques.”

The Claude Mythos Preview, a versatile model, can detect and exploit weaknesses “in all leading operating systems and web browsers when prompted,” according to Anthropic. 

Official usage rights are confined to select firms via the Project Glasswing initiative, involving companies like Nvidia, Google, Amazon Web Services, Apple, and Microsoft. 

Governments are also monitoring this innovation. Anthropic plans to withhold public release of the model due to fears of misuse.

An Anthropic representative stated to Bloomberg that they are assessing reports of unauthorised access to Claude Mythos Preview through one of their partner environments. 

So far, Anthropic has found no proof that this unauthorised access poses a risk to its systems or extends beyond the partner's setup.

The date of April 7 marks when Anthropic publicly announced limited access to Mythos for testing, coinciding with the reported unauthorised access. 

The group, linked to a Discord channel searching for info on undisclosed AI models, remains unnamed. 

Bloomberg details that the team gained entry using insights from Anthropic’s model designs, learned from a recent Mercor data breach, to pinpoint its location online.

This access has been actively used by the group, who shared screenshots and live demonstrations of the model with Bloomberg as evidence, while steering clear of using it for cybersecurity to evade detection by Anthropic. 

The group has also tapped into other unreleased Anthropic AI models, according to Bloomberg.