Anthropic says these topics are too dangerous to allow its version of Fable 5 to be discussed

Anthropic Tuesday Claude Fable 5 was released publiclyhis first “Mythos-class” model that he says surpasses his previous Opus models in all capabilities. But the launch of the brand today comes with safeguards designed to avoid answering questions about topics such as cybersecurity, biology, and chemistry, where the company has to openly complain about what might happen “Elevating” evildoers.

Anthropic says Fable 5 works on the “same model” as Mythos 5, which is coming out its time for the “Mythos Preview” months. today, but only for a “small group of cyberdefenders” judged to be reliable through The current Project Glasswing. Unlike Mythos 5, however, the publicly available Fable 5 is designed to support questions on some topics that are difficult to the previous version of Claude Opus 4.8 and warn the user when this happens.

Among the many things that have been said to have changed on Fable 5, those related to cybersecurity were the biggest jump.

Anthropic said it has designed these defenses to be “more robust than appropriate,” meaning that systems can sometimes reject “harmless requests” in a way that it admits can be frustrating for regular users. But Anthropic says that such false claims come up in less than five percent of all cases tested, and it was necessary to avoid situations in which Mythos would give aid to the perpetrators “to cause serious harm that they would not have received from other sources.”

I can’t let you do that, Dave

Fable 5’s head protection is built around system of classifiers designed to take serious note of the urgent bans and any attempts that may be made in the prison. In more than 1,000 hours of testing the red team and the bug bounty program, Anthropic says that external teams have failed to find the global prison of Fable 5. The new version also rejected the test of the independent prison to a greater extent than previous examples of Claude Opus, Anthropic said.

The company said it is particularly concerned about Mythos 5’s ability to perform “agentic hacking,” covering more cyberattacks and more locations than previous versions. But testing from the UK’s AI Security Institute in recent months found that Mythos Preview it did the same with OpenAI’s GPT-5.5 in the Capture the Flag challenge list, indicating that Mythos did “not win a single model.”