Anthropic · 2026-04-21 · major
Unauthorized Group Gains Access to Anthropic's Restricted Mythos AI Cybersecurity Model
On Mythos's launch day, a Discord community guessed its endpoint URL from Mercor breach data, then a third-party contractor shared access. The group gave Bloomberg a live demonstration — illustrating how contractor chains bypass even the most restricted AI access controls.
Anthropic's 'too dangerous to release' cybersecurity model was accessed by unauthorized users the day it was announced.
What is it?
Claude Mythos Preview is Anthropic's restricted AI model capable of autonomously finding and exploiting zero-day vulnerabilities. It was announced April 7, 2026 with access limited to ~50 vetted organizations via Project Glasswing.
How does it work?
A Discord community focused on accessing unreleased AI models used naming convention knowledge obtained from the Mercor data breach to guess Mythos's endpoint URL. An employee at a third-party contractor with authorized Glasswing access then shared credentials, allowing the group to access Mythos through shared API keys.
Why does it matter?
The group provided Bloomberg with screenshots and a live demonstration of Mythos access. Anthropic stated 'no evidence that its systems have been impacted' but confirmed the third-party vendor environment was compromised. The incident illustrates how even the most carefully controlled AI deployments are exposed through third-party contractor chains — and that a model explicitly framed as a dual-use cyberweapon risk had its controls bypassed within hours of launch.
Who is it for?
Security teams deploying frontier AI models via contractor networks should treat third-party vendor API key hygiene as a critical attack surface.