BODY: Anthropic's most tightly controlled AI model has fallen into the wrong hands, according to a new report from Bloomberg β raising fresh questions about whether restricted-access releases can actually contain powerful AI systems.
Claude Mythos Preview, announced by Anthropic on April 7, was never intended for the general public. The model demonstrated such significant cyberattack capabilities during internal testing that the company opted to limit access to a select group of vetted organizations β primarily cybersecurity firms and government-affiliated research bodies β rather than risk a wide release. The goal was to let defenders study the model's offensive potential and build countermeasures before any broader rollout.
Now, Bloomberg reports that unauthorized users have managed to obtain access credentials to Claude Mythos Preview. The details of exactly how access was compromised remain unclear, but the incident underscores a fundamental tension in AI safety strategy: models restricted precisely because of their dangerous capabilities become high-value targets for the very actors they're being shielded from.
Anthropic has not yet issued a detailed public response to the Bloomberg report. The company has previously emphasized its Responsible Scaling Policy as a framework for managing models that cross dangerous capability thresholds, and Claude Mythos Preview appears to be one of the first real-world stress tests of that framework.
The insider take
The incident is being closely watched in Tokyo's growing AI policy circles, where Japan's approach to AI governance has leaned toward enabling innovation rather than imposing hard restrictions. If a limited-access model can be breached this quickly, it strengthens the argument β already gaining traction among Japanese cybersecurity officials β that capability restrictions alone are insufficient without robust technical safeguards. For Anthropic, which has been deepening its ties with Japanese enterprise partners, the timing is particularly awkward: trust in access-control frameworks is exactly what large Japanese firms evaluate when choosing an AI vendor.
Originally reported by GIGAZINE (Japanese).