Anthropic briefs Pentagon on Claude Mythos: Could make large-scale cyberattacks far more likely

Anthropic is in active discussions with US government officials about its unreleased AI model Claude Mythos, after leaked internal documents revealed that the model possesses what the company itself describes as unprecedented cybersecurity risks.

Mythos, accidentally exposed through Anthropic's publishing system in late March, is the most powerful model the company has built. Internal assessments describe it as far ahead of any other AI model in cyber capabilities and warn that it could democratize advanced attack capabilities, making them accessible to threat actors with limited resources.

The model demonstrates dramatically improved performance in vulnerability discovery, exploit development, and multi-step attack reasoning. Experts warn that Mythos could industrialize cyberattacks by enabling vulnerability scanning and exploitation at a scale previously impossible.

The Department of Defense has expressed concern, particularly after Anthropic declined requests to allow the model for domestic surveillance or fully autonomous military weapons systems. Cybersecurity experts note that earlier Claude models were already used by a state-sponsored group to infiltrate various organizations, amplifying the concern around Mythos.

Anthropic plans a phased rollout limited to enterprise security teams and selected early access partners. The idea is that defenders get to understand and prepare for the model's offensive capabilities before potential attackers gain access. A public API, pricing, and confirmed release date have not been announced. The company describes the model as very expensive to serve.

Cybersecurity stocks fell following news of the leak. The paradox is stark: an AI model that could shift the global cybersecurity risk landscape, presented by a company that places safety at the forefront of everything it does.