Anthropic’s Claude Mythos has generated significant attention with claims of impressive capabilities. The model, which bears the name ‘Mythos,’ allegedly possesses remarkable abilities, including the capacity to find vulnerabilities and bugs across diverse applications, operating systems, web browsers, and legacy software.
Anthropic has presented this information through both a blog and a detailed 250-page report. The model is described with a ‘mythically-named’ nature, leading to its associated pitch. However, the text also notes discrepancies regarding the magnitude of its findings, pointing out that claims of identifying ‘thousands’ of severe zero-days rely on only 198 manual reviews.
Details within the documentation were reported to include over 20 pages where Anthropic staff wrote descriptions about their ‘novel impressions’ of the new model and its alleged ‘fondness for particular philosophers.’
In addition to the claims of deep technical capabilities, the promotional material also included repeated suggestions from Anthropic staff and the company itself concerning potential dangers. Specifically, the documentation contained repeated suggestions that concern should be raised, and even that people should be ‘terrified,’ of what AI like Claude Mythos can achieve.
Despite the high-stakes presentation regarding the model’s power, the text points out that Anthropic staff repeatedly suggest that they are unsure if this new AI model is conscious.
Regarding its distribution, Anthropic is concerned about the general release of the tool. Consequently, the company plans to keep Mythos internal, focusing instead on collaborations with major technology corporations and governments. This internal focus aims to prevent the tool from being misused, which could potentially cause untold mayhem.