Claude Mythos leaks as new Anthropic flagship
Anthropic's "Claude Mythos" model (codenamed Capybara) has been revealed through a major CMS leak, showing a "step change" in performance with estimated scores of 85-88% on SWE-bench Verified and 90%+ on MMLU. Positioned above the Opus line, Mythos is described as a 10-trillion parameter model with unprecedented cybersecurity capabilities currently undergoing private trials with early-access customers.
Mythos represents a shift toward a new tier of intelligence that could redefine AI's role in cyber-defense and autonomous engineering. Its reported performance suggests a model capable of autonomous codebase management, while its 10-trillion parameter scale signals a massive increase in compute intensity. However, the leak of 3,000 internal assets via a CMS misconfiguration highlights a significant security lapse for the safety-focused organization.
DISCOVERED
11d ago
2026-04-01
PUBLISHED
11d ago
2026-03-31
RELEVANCE
AUTHOR
assymetry1