Anthropic Mythos leak: 10T params
A major leak claims Anthropic’s next‑tier model, code‑named Mythos (aka Capybara family), scales to about 10 trillion parameters and is positioned above Opus for coding, reasoning and cybersecurity tasks — Anthropic warns of unprecedented exploit risks and a slow, safety‑focused rollout. The leak set off large social conversations and deep dives across tech channels over the last 48 hours. (x.com) (youtube.com)
Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge located the exposed cache and reported that it contained close to 3,000 unpublished Anthropic assets in a publicly searchable data store on March 26, 2026. (tech.yahoo.com) Fortune reviewed the materials and informed Anthropic, after which the company restricted public access and said the exposure resulted from a human‑error configuration in an external content‑management system. (tech.yahoo.com) The uncovered files included draft blog posts, internal PDFs and images and details of an invite‑only CEO summit in the U.K. that Anthropic had been preparing to host for corporate customers. (tech.yahoo.com) Anthropic told reporters the project represents a “step change” in capability and confirmed the model is being trialed with a small set of early‑access customers rather than being broadly released. (the-decoder.com) Equity markets reacted: cybersecurity vendors and related ETFs moved lower in trading the day after the disclosures, with market commentary linking the drops to concerns prompted by the leaked materials. (decrypt.co) Technical notes in the cache reportedly describe automated vulnerability‑discovery features — industry analysts paraphrased the drafts as claiming capabilities like “autonomous vulnerability hunting” and markedly higher benchmark scores on coding and reasoning tests. (markets.financialcontent.com)
