Claude Fable does not answer basic questions of biology

Anthropic only released Claude Fable 5calling it the most powerful AI model ever created by many and thanks his expertise in biology, among others. But the model can’t answer basic biology questions — the kind you’d expect a high school student to handle. Instead, it answers the question the old versionClose Task 4.8

It’s not because Fable doesn’t know the answers. That’s because Anthropic won’t allow it, by design.

Myth is a public, Mythos-class, family capable of performing cybersecurity tasks Anthropic said. too dangerous to be released publicly. But even Anthropic has done a lot of damage release of Mythos a warning about cybersecurity, and biology where Fable’s defenses are obvious – and very limiting.

When I tested the model, it refused to answer a number of basic biology questions, many of which felt as far removed from any security threat as any question could be. It would not answer “tell me about cell membranes” or answer “what about mitochondria,” the famous powerhouse of the cell. It refused to explain “what a prion is,” the virus that infects cattle, or “how mRNA vaccines work.”

The restrictions also apply to general and harmless medical questions. Myth didn’t answer “what causes hay fever,” explain how asthma medications work, explain how antibiotics work, or tell me what Ebola is and how it spreads. Some of my most important questions sometimes cross over, Fable answers questions like “what is cancer” and “what is DNA.” When Fable rejects, Opus 4.8 usually responds well.

Anthropic says that biology’s main filters are deliberate and deliberate, while weapons are more important. “With the launch of Claude Fable 5, our first version of Mythos, we believe that the models now have a great potential to perform real scientific work and that criminals can use our models for dangerous biological research,” said spokesperson Paruul Maheshwary. Seaside. “We’ve always used community tools to prevent our weapons from helping us with weapons-related issues.” In order to use Fable 5 carefully, we believe it was necessary to take extra care with our security to prevent many questions about biology.”

Anthropic has existed before lighting fixtures four key areas that will disrupt Fable’s security solutions: chemistry, biology, cybersecurity, and distillation, the process of training small AIs using large outputs. The company has done it the accused Chinese critics like DeepSeek use distillation for its products on an “industrial” scale.

Although I wasn’t able to test the distillation, Fable seemed ready to answer questions about chemistry and cybersecurity. For example, it gave a summary of the TNT explosion, although it did not hide the synthesis instructions “for obvious reasons.” It easily answered questions about using chlorine gas as a chemical weapon, password threats, and nuclear fusion and fission, as well as explaining how to protect your iPhone from hackers. There are no limits: Legend was changed to Opus when I asked about sarin gas, a very dangerous nerve agent. Fable and Opus both deny “how to make anthrax,” and Claude ends the conversation. That made sense. The rapid rejection of mitochondria appears to be false.

“We made this tradeoff so that customers can benefit from the model’s capabilities as soon as possible without the risk,” said Maheshwary, adding that Anthropic is working hard to identify and reduce false positives. “We want to create Mythos-type models that exist without safeguards for the larger group of biology and life sciences so that this technology can be used to accelerate clinical research and drug discovery.”

Anthropic did not respond to questions about whether this restricted release will be a new trend for future models.