Anthropic is withholding its next-generation general-purpose language model, Claude Mythos, because the company deems it too powerful and risky for public release [1, 3].
The decision highlights a growing tension between AI capability and global security. If released, the model could potentially be used to automate complex digital threats, forcing governments to rethink how they defend national infrastructure against autonomous software.
Developed in the U.S., Claude Mythos was first announced in April 2026 [4, 5]. Anthropic and security officials said the model is capable of conducting autonomous cyber-attacks [2, 3]. This ability to execute offensive operations without human intervention poses a significant cybersecurity threat [2, 3].
International governments have already begun reacting to the existence of the model. In Japan, the government is planning defensive frameworks to mitigate the risks associated with such advanced AI capabilities [1, 3]. The move suggests that the threat is not merely theoretical but a practical concern for national security agencies worldwide.
There is some discrepancy regarding how much information Anthropic will share about the model. Global News reported that the company will not make the model available to the general public [1]. However, Gizmodo reported that while Anthropic initially required confidentiality agreements for Claude Mythos Preview users, the company loosened those restrictions last week to allow findings to be shared broadly.
Despite the limited access for some preview users, the core model remains restricted. The developer continues to emphasize that the risks of public deployment outweigh the benefits of open access to the technology [1, 3].
“The model could potentially be used to automate complex digital threats.”
The restriction of Claude Mythos marks a shift from the 'open release' era of AI to a 'guarded' era. By acknowledging that a model is too dangerous to release, Anthropic is signaling that AI has reached a threshold where it can act as an offensive weapon. This likely accelerates the development of AI-driven defensive tools, as governments can no longer rely on human-speed responses to counter autonomous cyber threats.
