Anthropic's Mythos model is good enough at finding software vulnerabilities that the company won't hand it to just anyone. So it handed it to 200 organizations instead.
In early June, Anthropic expanded access to Mythos to 150 additional organizations, bringing the total approved list to roughly 200 across 15 countries. The model is described as capable enough at hunting software vulnerabilities to enable data theft or disruption of critical infrastructure in the wrong hands. Rather than shelve it or open it broadly, Anthropic has settled on controlled access: a vetted set of partners operating under conditions presumably stricter than general availability.
The arrangement puts a concrete number on what "too dangerous for the public" actually means: about 200 organizations, spread across a portion of the globe. That framing matters, because Anthropic is effectively arguing that risk is a function of who holds the tool, not whether the tool exists at all. If the model grows more capable, or the approved list keeps expanding, that argument gets harder to sustain.
Anthropicisn't the first lab to describe its own technology as dangerous while continuing to distribute it. It probably won't be the last.