Leak: Anthropic’s ‘Mythos’
A leak says Anthropic is preparing ‘Mythos,’ a compute‑heavy LLM aimed at cybersecurity use cases and initially offered to enterprise security teams. The reported specialization suggests models will be verticalized for high‑stakes domains, with limited early access. (csoonline.com)
Fortune’s March 26 reporting found roughly 3,000 unpublished Anthropic assets sitting in a publicly searchable content cache after security researchers Roy Paz (LayerX) and Alexandre Pauwels (University of Cambridge) located the files. (fortune.com) Anthropic told reporters the exposed drafts describe a model it regards as a “step change” and “the most capable we’ve built to date,” and said the system is already being trialed with a small set of early‑access customers. (tech.yahoo.com) Leaked drafts name a new top tier internally called “Capybara” above the Opus line and claim Mythos/Cabybara posts “dramatically higher” benchmark scores in software coding, multi‑step academic reasoning, and cybersecurity tasks versus Opus 4.6. (coindesk.com) Anthropic’s own notes in the exposed material state Mythos is “very expensive for us to serve, and will be very expensive for our customers to use,” and the company says it plans a deliberately slow, security‑focused rollout while it works to improve efficiency before any general release. (computerworld.com) Markets reacted immediately: cybersecurity names including CrowdStrike and Palo Alto Networks fell roughly 6–7% on March 27, and the Global X Cybersecurity ETF dropped about 4.5% as investors priced in the risk of more capable AI tools shifting attacker/defender dynamics. (247wallst.com) Separately, The Information and Bloomberg reported Anthropic has held banker talks about a possible fourth‑quarter 2026 IPO that could raise north of $60 billion, following a private valuation around $380 billion in February’s funding round. (theinformation.com)
