Anthropic's restricted Mythos cybersecurity model leaks the day it is announced
What Happened
Bloomberg reported on April 21 that a small group of users in a private Discord channel got hold of Anthropic's Mythos preview, a model the company says can identify thousands of OS and browser vulnerabilities. Anthropic had restricted Mythos to vetted security partners under "Project Glasswing." The unauthorized access surfaced the same day as the official announcement.
My Take
Treating this purely as a leak misses the structural problem. Any model behind a "trust us" perimeter is one disgruntled tester away from being public, and the security community has known this since GPT-4. Anthropic shipped the announcement before access controls were hardened — presumably because the PR value of being first dropped fast as competitors closed in. Prediction: within a quarter, major labs adopt audited enclave standards for restricted models, modeled on cleared compartments rather than Discord invites.
Read Original Source