Anthropic built Mythos, then refused to release it over security risks

Anthropic has declined to release its new Mythos Preview model publicly, citing its exceptional danger. According to the company, the model is so good at finding and exploiting software vulnerabilities that a public release would threaten economic stability, public safety and national interests. It is a rare case of a major AI company choosing to hold back its own product, and it has already sparked debate among industry specialists.

Mythos Preview is an experimental language model from Anthropic, which the company describes as too powerful for public release. Available information suggests the model shows outstanding cybersecurity capabilities: it can independently discover vulnerabilities in code and build chains to exploit them. In practice, this means Mythos could work as a highly skilled hacker-analyst, quickly, at scale and without human oversight.

Anthropic has not published technical details about the model's training or any independent safety evaluations. It has only promised to give access to Mythos Preview to a select group of researchers and government bodies.

The decision has drawn a mixed reaction from experts. Some AI safety researchers agree that powerful offensive cybersecurity tools do deserve special caution when distributed. Attackers already make heavy use of existing automated vulnerability scanners, and an LLM with advanced reasoning could take that threat to a fundamentally new level.

Other experts are skeptical. Anthropic has provided no independent evaluations or reproducible tests of Mythos's capabilities. Moreover, restricting access indirectly benefits the company itself: it looks like a responsible industry leader without risking public criticism of the model. Observers also note that these claims coincide with active debates about AI regulation in the US and EU governments.

Skeptics are asking whether this is a real threat or a skillfully built PR campaign. Anthropic has consistently positioned itself as the safety-focused AI company, unlike competitors that it implies put release speed ahead of caution. That positioning attracts investors, regulatory goodwill and the trust of enterprise customers. The Mythos decision fits this logic neatly: the company shows it is willing to sacrifice commercial access for the sake of safety.

But critics point out that you cannot be judge and defendant at once. No independent assessment of the model's capabilities exists, only Anthropic's word.

The Mythos story raises a question that will be central for the whole industry in the coming years: who decides how dangerous a given AI system is, and who checks that the decision is right? If companies draw that line themselves, a conflict of interest is unavoidable. The Mythos Preview case may become the first real test case for the regulatory mechanisms now being developed on both sides of the Atlantic.