Leaked 'Claude Mythos' docs surface
What happened
Leaked Anthropic documents describing a new 'Claude Mythos' 'Capybara' tier—purportedly stronger at coding, reasoning, and cybersecurity—have circulated on social, and markets reacted with cyber‑security stocks falling 4–6% while BTC moved to about $66K. Anthropic reportedly offered defenders early access ahead of any broad release. (x.com) (x.com)
Why it matters
Security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) found a publicly accessible content cache linked to Anthropic’s blog that contained nearly 3,000 unpublished assets, and Fortune reviewed and alerted Anthropic on March 26, 2026. (fortune.com) A draft blog in the exposed cache named an internal project “Claude Mythos,” referred to an internal tier codenamed “Capybara,” and used language characterizing the work as a “step change” and the company’s most capable system to date, which Anthropic acknowledged after Fortune’s inquiry. (fortune.com) The leaked drafts explicitly warn the model is “currently far ahead of any other AI model in cyber capabilities” and detail automated vulnerability-discovery and exploitation workflows that security analysts cited as the proximate cause of market selling in software names. (techzine.eu) U.S. cybersecurity equities opened sharply lower after the disclosures, with CrowdStrike down about 7%, Palo Alto Networks off roughly 6%, Zscaler around 4.5% and Okta, SentinelOne and Fortinet each near 3% losses as trading began on March 27. (finance.yahoo.com) The draft materials also flagged operational constraints—describing the system as “very expensive to serve” and outlining a staged, security‑sensitive rollout—and the exposed cache included other unpublished corporate items such as an invite‑only CEO summit agenda. (ai-primer.com) Anthropic’s public red‑team work since March 2025 and the company’s November 2025 disclosure about an AI‑driven espionage campaign have already driven corporate security reprioritization, and analysts are pointing to the Mythos leak as a catalyst for accelerated hiring of red‑teamers, vulnerability researchers and AI safety/security engineers. (anthropic.com)
Key numbers
- Leaked Anthropic documents describing a new 'Claude Mythos' 'Capybara' tier—purportedly stronger at coding, reasoning, and cybersecurity—have circulated on social, and markets reacted with cyber‑security stocks falling 4–6% while BTC moved to about $66K.
- cybersecurity equities opened sharply lower after the disclosures, with CrowdStrike down about 7%, Palo Alto Networks off roughly 6%, Zscaler around 4.5% and Okta, SentinelOne and Fortinet each near 3% losses as trading began on March 27.
Quick answers
What happened in Leaked 'Claude Mythos' docs surface?
Leaked Anthropic documents describing a new 'Claude Mythos' 'Capybara' tier—purportedly stronger at coding, reasoning, and cybersecurity—have circulated on social, and markets reacted with cyber‑security stocks falling 4–6% while BTC moved to about $66K. Anthropic reportedly offered defenders early access ahead of any broad release. (x.com) (x.com)
Why does Leaked 'Claude Mythos' docs surface matter?
Security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) found a publicly accessible content cache linked to Anthropic’s blog that contained nearly 3,000 unpublished assets, and Fortune reviewed and alerted Anthropic on March 26, 2026. (fortune.com) A draft blog in the exposed cache named an internal project “Claude Mythos,” referred to an internal tier codenamed “Capybara,” and used language characterizing the work as a “step change” and the company’s most capable system to date, which Anthropic acknowledged after Fortune’s inquiry. (fortune.com) The leaked drafts explicitly warn the model is “currently far ahead of any other AI model in cyber capabilities” and detail automated vulnerability-discovery and exploitation workflows that security analysts cited as the proximate cause of market selling in software names. (techzine.eu) U.S. cybersecurity equities opened sharply lower after the disclosures, with CrowdStrike down about 7%, Palo Alto Networks off roughly 6%, Zscaler around 4.5% and Okta, SentinelOne and Fortinet each near 3% losses as trading began on March 27. (finance.yahoo.com) The draft materials also flagged operational constraints—describing the system as “very expensive to serve” and outlining a staged, security‑sensitive rollout—and the exposed cache included other unpublished corporate items such as an invite‑only CEO summit agenda. (ai-primer.com) Anthropic’s public red‑team work since March 2025 and the company’s November 2025 disclosure about an AI‑driven espionage campaign have already driven corporate security reprioritization, and analysts are pointing to the Mythos leak as a catalyst for accelerated hiring of red‑teamers, vulnerability researchers and AI safety/security engineers. (anthropic.com)
