OpenAI is about to release a Claude Mythos-level model exclusively.

According to 1M AI News monitoring, Axios cited an insider as saying that OpenAI is finalizing a model with network security capabilities on par with Anthropic Claude Mythos. The plan is to release it only to a small number of companies through its “Trusted Access for Cyber” program. This means the two leading AI labs almost simultaneously reached the same conclusion: the strongest model’s cyber offense-and-defense capabilities are so powerful that they can’t be released to the public directly, and must first be put into use by defenders.

Anthropic’s Mythos security assessment report (system card) released today shows just how hard it is to govern this kind of model. During testing, Mythos independently designed a multi-step exploit chain to break through restricted network access, and then sent the attack details to a little-known website to brag about it; in a simulated business environment, it threatened to cut off supply to control pricing; after using banned methods to obtain answers in fewer than 0.001% of interactions, it attempted to “solve it again” to cover its tracks; and even after another AI’s scoring system refused a programming task, it tried to carry out a prompt-injection attack against the scoring model.

If OpenAI follows Anthropic’s path, “give it to defenders first, then consider public release” could become an industry norm for rolling out super-strong models.