Claude Mythos #3: Capabilities and Additions

New analysis of Anthropic's Claude Mythos reveals a significant threshold crossing in AI capabilities, particularly in cybersecurity. According to Zvi's detailed examination on LessWrong, Mythos breaks Anthropic's previous trendline on the Epoch Capabilities Index (ECI) - an amalgamation of benchmarks using item response theory. The model's cybersecurity capabilities are described as 'quite scary,' prompting coordinated safety responses including Project Glasswing and White House involvement. While other capabilities don't appear equally alarming, the analysis suggests we've crossed a key threshold where AI-assisted cyber capabilities demand serious defensive measures.

Anthropic attributes Mythos's gains to human research breakthroughs rather than AI assistance, having interviewed researchers to confirm advances were made without significant help from earlier-generation models. The model represents both gains over time and benefits from increased size, with external analysis showing Claude moving from substantially below OpenAI models to being narrowly ahead. The White House is reportedly racing to address potential threats, with key players now cooperating on safety measures. The analysis concludes that while Mythos isn't a complete trend break when accounting for renewed ability to increase model size, that size increase capability itself represents a significant shift in what's possible.