Contractor links leak unreleased models

Unauthorized users gained access to Anthropic’s unreleased Mythos model by combining contractor‑granted permissions, publicly exposed artifacts (GitHub, breach data), and online sleuthing in private channels. The incident shows that unreleased model locations and access can be inferred and misused even without direct compromise of vendor production systems. — Highlights a recurring governance and security gap: third‑party contractor credentials plus public provenance leaks create an emergent vector for leaking powerful unreleased AI systems.