Claude Mythos went from internal Anthropic codename to controlled public launch in roughly six months. This page walks through the public milestones (the early leaks, the accidental exposure on March 26 2026, the formal April 8 announcement, the AISI evaluation and the Project Glasswing onboarding waves), with sources for each entry.
Updated: May 9, 2026
Anthropic engineers internally referenced a new model tier above Opus under the codename "Capybara" through Q4 2025. The codename surfaced via job postings and conference talks that mentioned a tier classification beyond the Opus naming family. Public attention was minimal at the time.
A misconfigured Anthropic load balancer briefly exposed a model endpoint labelled "claude-mythos-preview" on the public API gateway. Several developers logged calls before the endpoint was withdrawn within 90 minutes. The leaked outputs included unusually long chain-of-thought traces with unexplained reasoning depth, leading to the first wave of public speculation about a Capybara-tier model.
On April 7, the FreeBSD security team published an emergency advisory for an NFS server RCE that would later be assigned CVE-2026-4747. The advisory thanked Anthropic for coordinated disclosure but did not name the discovery method. The patch was tagged 24 hours before the public Mythos announcement.
Anthropic published the Mythos preview announcement, the 244-page system card, and the launch of Project Glasswing on the same day. The announcement explicitly stated that Mythos would not be made generally available because of its cybersecurity capabilities. Twelve founding partners were named on day one (Amazon, Apple, Google, Microsoft and eight unnamed critical-infrastructure operators).
The UK AI Safety Institute (AISI) published its independent evaluation of Mythos's cyber capabilities, broadly confirming Anthropic's reported numbers and adding a separate offensive-capability assessment. The report concluded that controlled access via Glasswing was a reasonable mitigation given current alignment evaluation limits.
Anthropic began onboarding the expansion cohort of ~28 additional Glasswing partners through May 2026: Tier-1 telecoms, financial market operators and selected open-source maintainers via the 'Claude for Open Source' lane. The total programme size reached ~40 organisations by mid-May.
Anthropic stated it did not; the announcement schedule was already set for April 8. The leak compressed the press cycle but the technical and contractual timeline was unchanged.
April 8 2027. Anthropic has not yet committed to what happens after; the most likely outcomes are an extension as a permanent restricted-access programme or a graduated public release with hardened mitigations.