Anthropic created the Mythos model — and chose not to release it over security risks

Anthropic refused to release the new Mythos Preview model to public access, citing its exceptional danger. According to the company, the model is so proficient at finding and exploiting software vulnerabilities that its public release would pose a threat to economic stability, public safety, and national interests. This is a rare case where a major AI company decides to withhold its own development — and it has already sparked disputes among industry experts.

Mythos Preview is an experimental language model from Anthropic, which the company characterized as too powerful for public release. According to available information, the model demonstrates outstanding capabilities in cybersecurity: it can independently detect vulnerabilities in code and construct chains of their exploitation. In essence, this means that Mythos can operate as a highly qualified hacker-analyst — quickly, scalably, and without human oversight.

Anthropic has not published technical details about the model's training process or independent security assessments, limiting itself to promising access to Mythos Preview only to select researchers and government agencies. The company's decision sparked mixed reactions in the expert community. Some AI safety researchers acknowledge that powerful offensive cybersecurity tools do deserve special caution in distribution.

Existing automated vulnerability scanners are already actively used by malicious actors — and LLMs with advanced reasoning capability could raise this threat to a fundamentally new level. Another part of experts is skeptical. Anthropic did not provide independent assessments or reproducible tests of Mythos's capabilities.

Moreover, restricting access indirectly benefits the company itself: it looks like a responsible industry leader without bearing the risks of public criticism of the model. Observers also point out that such statements coincide with active discussions about AI regulation in U.S.

and EU governments. The question skeptics ask: is this a real threat or skillfully constructed PR? Anthropic consistently positions itself as a safe AI company — unlike competitors, allegedly prioritizing release speed over caution.

This positioning attracts investors, regulator loyalty, and trust from corporate clients. The decision on Mythos perfectly fits this logic: the company demonstrates willingness to sacrifice commercial access for safety. But critics point out: you cannot be both judge and defendant.

There is no independent assessment of the model's capabilities — there is only Anthropic's word. The Mythos story raises a question that will become key for the entire industry in the coming years: who exactly decides how dangerous a particular AI system is, and who verifies the correctness of this decision? If companies set this boundary themselves, a conflict of interest is inevitable.

The Mythos Preview case could become the first real test case for the regulatory mechanisms currently being developed on both sides of the Atlantic.