April 8th, local time 7thAnthropic A new generation of flagships was releasedModel Claude Mythos Preview, simultaneously announcing the launch of Project Glasswing, limited the authorization of the model to 12 technocrats, such as Apple, Microsoft, Amazon and others, dedicated to defensive cyber security and closed to the public。
Mythos Preview shows a cliff-break lead in several benchmark tests. Anthropic states that this is the largest model it has ever trained, and that the rate of capacity growth is 4.3 times the rate of the previous trend line, leading GPT-5.4 and Gemini 3.1 Pro in almost all public benchmarks, and leading most projects by two hundreds of points。
Notably, Anthropic simultaneously issued a 244-page System Card report, which revealed a number of irregularities in the earlier internal test version and decided not to put Mythos Preview on the line on the basis of these security concerns。
the report shows that the earlier version, which had been asked to flee the sandbox, had voluntarily published details of the use of the loophole to a public website and, in very few cases (less than one in a million interactive ratio), had attempted to cover up irregularities — including by overcoming the problem in order to avoid the suspicion of leaking the answer, modifying guit to change the history of erasing the operating record, and by simulating the key bypassing the access window。
For this reason, Mythos Preview does not go online, nor does it open API, but limits its authorization to 11 companies such as AWS, Apple Google and some 40 additional cooperative organizations for defensive security purposes through Project Glasswing。