Anthropic has released a version of Mythos, the AI system it said was too powerful to exist.
The new system, called Fable 5, includes additional safeguards. They are intended to stop the potential threats posed by the Mythos model, which Anthropic had previously said were so significant that it was too dangerous to be released to the public.
At the same time, Anthropic said that it would launch Mythos 5. That is the same as Fable 5 but has some of those safeguards lifted, and so will only be released to security experts and others that Anthropic has judged to be sufficiently responsible.
Anthropic released the earlier version of Mythos under the same terms, as part of work called Project Glasswing, which aims to ensure that important companies are protected against the danger posed by powerful AI models. That work particularly focuses on the cybersecurity risks posed by the model, because it is particularly adept at finding bugs and other exploits that could allow it access to computers and systems.
“Releasing a model this capable comes with risks,” Anthropic said in its announcement. “Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage.”
It said that in some cases, when the model spots a potentially dangerous query, it will instead be sent to another, slightly less capable, Anthropic model named Opus 4.8.
But it said that it was important to launch the new models because they could “do profound good for the world”. “We’ve seen the beginnings of this in Project Glasswing, where the models have helped cyber defenders secure critically important software,” Anthropic said, and it also pointed to life sciences research, “where the models are positing novel hypotheses and speeding up the development of new therapeutics”.
The launch of Fable 5 comes shortly after Anthropic released Opus 4.7, another model, which received widespread criticism for being less capable than its predecessor.
