Anthropic's Mythos AI model accessed by unauthorized group via contractor breach

Anthropic's most powerful cybersecurity AI model, Mythos, has been compromised after a small group of unauthorized users gained access through a third-party contractor. The incident, first reported by Bloomberg, raises serious concerns about the security of advanced AI systems that the company itself has described as potentially dangerous in the wrong hands.

The breach involved a private online forum where members used a mix of tactics, including exploiting a contractor's access credentials and employing commonly used internet sleuthing tools. The contractor, who remains unnamed, reportedly facilitated the entry point. Mythos, officially called the Claude Mythos Preview, is a general-purpose model capable of identifying and exploiting vulnerabilities across major operating systems and web browsers. Anthropic had previously warned that such capabilities could pose significant risks if misused.

The incident highlights growing pressure on AI developers to secure their most advanced models against insider threats and external exploitation. While the full extent of the breach remains unclear, the unauthorized access signals potential vulnerabilities in Anthropic's access controls and contractor management. The company now faces scrutiny over how it safeguards models that could be weaponized for cyberattacks, with implications for the broader AI industry's approach to security and responsible deployment.