Anthropic confirms 'Claude Mythos' test
Anthropic confirmed it’s testing a new, more powerful model dubbed “Claude Mythos” after a leak exposed details — the company has throttled Claude access during peak hours while it investigates security and stability risks, and the episode sent related stocks sliding (notably Cloudflare). ( )
Documents were found in an unsecured, publicly searchable data lake after a CMS misconfiguration exposed the cache on March 26, 2026. (tech.yahoo.com)) The exposed repository contained nearly 3,000 internal CMS assets, including draft blog posts and product materials, according to reporting that reviewed the cache. (benzinga.com)) Leaked drafts referenced a higher-performance tier codenamed “Capybara” above the Opus tier and included logistical notes about an invite‑only CEO summit in Europe aimed at enterprise customers. (livemint.com)) Internal notes in the cache warned the model could identify and exploit software vulnerabilities at “machine speed,” framing Mythos as a capability jump that posed novel cybersecurity risks. (markets.financialcontent.com)) LayerX security researcher Roy Paz and University of Cambridge researcher Alexandre Pauwels publicly located and signaled the presence of the unsecured materials during independent reviews. (tech.yahoo.com)) Market moves followed the disclosures on March 27, 2026: CrowdStrike closed down about 5.9%, Palo Alto Networks fell roughly 6.0%, and Zscaler dropped near 5.9%, while Cloudflare shares closed near $203 amid the sector sell‑off. (financecharts.com)) Reporting from multiple outlets noted the draft language described Mythos as a “step change” in capability and the leaked materials prompted fresh scrutiny of how Anthropic stored early‑stage communications and product drafts. (mashable.com))
