Anthropic's AI Mythos: Companies still have access to preview version

Despite the US government's order to block foreign access to Anthropic's AI models Fable and Mythos, the preview version is apparently still available to various companies. Some companies told the US business news service Bloomberg that they still have access to preview versions as part of the "Glasswing" project.

The Mythos preview had been made available to several companies as part of an initiative called Project Glasswing, primarily in the IT and finance sectors. Anthropic refrained from a general release of Mythos because the AI model was too dangerous. A Mythos variant named Fable 5, classified as safe by Anthropic, was briefly available but had to be taken offline again following a US government order. According to Anthropic, the order included Mythos 5 and Fable 5. However, as part of Project Glasswing, companies likely had access to earlier Mythos versions that were not affected.

Companies confirm access to Mythos

Several companies told Bloomberg that they still have access to a preview version older than Mythos 5. Only the IT security consultancy Dragos and network specialist Cisco officially confirmed this. Anthropic declined to comment to Bloomberg. According to Bloomberg, other participants in Project Glasswing lost access after the US order, such as the EU cybersecurity agency Enisa.

Meanwhile, international US companies using Anthropic's AI models are implementing strict geographical rules. For example, the US bank JPMorgan Chase has cut off access for its staff in Hong Kong, reports the Financial Times (FT). It follows US bank Goldman Sachs. In both cases, Anthropic's terms of use are said to be the reason. In the case of Goldman Sachs, it is known that Anthropic's license agreement prohibits use in the Greater China region.

The trigger for the sudden blocking of Fable 5 and Mythos 5 was a tip from Amazon to the US government that it had been possible to bypass Fable 5's strict security precautions with a trick – according to Amazon, as part of a so-called "jailbreak." Amazon's report on this is not public. This week, an IT security researcher spoke out, claiming to have seen the Amazon report. In her view, it is not a jailbreak, and the US government's order was not proportionate.

Meanwhile, the US government accuses Anthropic of allowing China to access Mythos. Reports of unauthorized access to Mythos in general had already emerged shortly after Mythos was introduced.

(nen)