Anthropic's dangerous AI Mythos: Unauthorized access likely since day one

Anthropic's most powerful AI is said to be so dangerous that only selected companies have access. Unknown individuals reportedly bypassed this, a report states.

A group of individuals has allegedly gained unauthorized access to Anthropic's powerful and downright dangerous AI model Claude Mythos Preview without the AI company noticing. This is reported by the financial news agency Bloomberg, which was shown the use of the tool. The unknown individuals are reportedly communicating in a private Discord channel and are individuals who have previously focused on searching for unpublished AI models. Mythos was not used by them for tasks related to cybersecurity; instead, they are testing how the AI model performs on harmless tasks - for example, building a website.

Anthropic introduced Mythos two weeks ago and stated that the model is so dangerous that it is only made available to companies working on IT security. The AI model has already identified thousands of high-risk zero-day vulnerabilities, including in all major operating systems and every internet browser. At the same time, the AI technology is significantly more capable of developing a working exploit for such vulnerabilities, sometimes even using multiple in conjunction. As part of "Project Glasswing", the industry is now to work on patching vulnerabilities found this way before other AI models become available, with which criminals can also find and, above all, exploit vulnerabilities much more easily.

Unauthorized access to Mythos was obtained according to Bloomberg on the very day Anthropic introduced the AI tool. The group used various tactics, with one person posing as an employee of a service provider to gain access to Anthropic's tools. Previously, the group had made an "educated guess" about Mythos's internet address - based on other Anthropic URLs. The unknown individuals have been using Mythos regularly since then, just like other AI models before. However, their intention is not "not wreaking havoc with them". Bloomberg consistently refers to one of the individuals for the report, who is kept anonymous.

Anthropic has reportedly pledged to investigate the claim, while downplaying the extent of the access. There are no indications that the access went beyond a third-party environment or had any impact on its systems. The discovery suggests how difficult it may be for the company to keep access to Mythos under wraps. The AI model was described at its introduction as so powerful that it not only alarmed the IT security industry. Governments in more and more countries are grappling with the significance of the new tool, and checks have been ordered, especially in the financial industry. If Mythos falls into the wrong hands, the consequences for cybersecurity could be devastating.