'Claude Mythos': Anthropic confirms testing its most advanced model yet
What's the story
Anthropic has confirmed to Fortune that it is testing its most advanced AI model yet. The announcement comes after a data leak (reviewed by publication) revealed details about the upcoming model, dubbed "Claude Mythos." While this new AI is expected to be a major leap forward from its predecessors, Anthropic has also warned that it could pose significant cybersecurity risks.
Information breach
Leak revealed details about new AI model
The leak, which originated from Anthropic's content management system (CMS), exposed nearly 3,000 unpublished assets. These included draft blog posts, images, and PDFs related to the upcoming AI model. The leaked information revealed that Claude Mythos would feature a new tier called Capybara. This tier is bigger and better than Anthropic's previous top models, Opus, and is expected to perform significantly better on reasoning, coding, and cybersecurity tests.
Risk assessment
Claude Mythos could trigger new cyberattacks
A draft blog post from Anthropic warned that Claude Mythos is "currently far ahead of any other AI model in cyber capabilities" and could trigger a new wave of cyberattacks that outpace current defense efforts. To counter this, the company plans to release the model initially to cyber defenders. This would help them fortify codebases against potential AI-driven exploits.
Security measures
Plans for exclusive retreat to discuss AI adoption leaked
Anthropic has previously thwarted hacking attempts by groups linked to the Chinese government. One such campaign targeted multiple organizations across technology, finance, and government sectors. The leaked documents also revealed plans for an invite-only retreat for European business leaders at an 18th-century manor in the English countryside. Anthropic CEO Dario Amodei is expected to attend this event where discussions on AI adoption and demonstrations of unreleased Claude capabilities will take place.