Anthropic Probes Unauthorized Access to Restricted AI Model

Anthropic is investigating reports that a small group of unauthorized users gained access to Claude Mythos, its restricted AI model capable of enabling dangerous cyberattacks. The breach reportedly occurred through credentials belonging to a third-party contractor, combined with data from a breach at AI startup Mercor Inc. Anthropic says there is no indication its own systems were compromised beyond the vendor environment. The group has reportedly avoided offensive tasks to evade detection, but security experts warn the incident exposes critical weaknesses in supply chain access controls and highlights growing risks from AI-enabled adversaries.