Claude Mythos leak surfaces
A major social-media leak claims Anthropic’s new 'Claude Mythos' model details were exposed, sparking fresh debate about model safety and cybersecurity risks for new pretrained systems. The discussion on X highlights how quickly operational and safety questions spread when proprietary model artifacts leak. (x.com)
Fortune published the first report on March 26, 2026 that unpublished Anthropic content revealed an unreleased model called Claude Mythos and prompted company acknowledgement. (fortune.com) A content‑management misconfiguration left nearly 3,000 unpublished Anthropic assets publicly discoverable before access was locked down, according to reporting on the exposed data cache. (winbuzzer.com) Anthropic told reporters the company has completed training and is testing the new model, describing Mythos as a “step change” and “the most capable we’ve built to date” in draft materials. (tech.yahoo.com) Leaked draft copy warned Mythos is “currently far ahead of any other AI model in cyber capabilities” and could identify and exploit software vulnerabilities faster than defenders can respond, per the exposed blog draft. (decrypt.co) Internal documents surfaced in the cache pointed to an internal codename, Capybara, for a new model tier positioned above the existing Opus lineup and aimed at enterprise and government use. (ctol.digital) Anthropic’s drafts and assets included unpublished blog drafts, PDFs and event planning notes for invite‑only customer outreach, all of which were among the files exposed. (mashable.com) The leak triggered immediate market moves: reports showed a sector selloff on March 27, 2026 with major cybersecurity names sliding amid investor concern about Mythos’s offensive capabilities. (marketwatch.com) Anthropic attributed the exposure to human error in a CMS configuration, restricted access after discovery, and said the materials were early drafts rather than a public release. (tech.yahoo.com)
