Anthropic announced the launch of Claude Fable 5 this week, touting it as the company’s most capable model to date and the first Mythos‑class AI made widely accessible. The announcement sparked excitement among developers and researchers eager to test its promised scientific prowess. Within hours, however, users discovered a puzzling limitation: the model would not answer many straightforward biology questions.

When prompted with basics such as “What are cell membranes?” or “Explain mitochondria,” Claude Fable 5 either refused outright or handed the request to its predecessor, Claude Opus 4.8. The older model, still powerful, supplied the answers without hesitation. The same pattern emerged for queries about prions, the mechanism of mRNA vaccines, hay fever, asthma medication, antibiotic resistance and even the nature of Ebola. Only broader topics—like “What is DNA?” or “What is cancer?”—received a response from Fable.

Anthropic’s spokesperson, Paruul Maheshwary, explained that the behavior is by design. “We made this trade‑off so customers could benefit from the model’s capabilities sooner without the risks,” Maheshwary told The Verge. The company has long flagged bioweapon research as a high‑risk use case for large language models. To mitigate that threat, Anthropic applied an “overly conservative” filter to any request touching on biology, chemistry, cybersecurity or model distillation.

In practice, the filter blocks most biology‑related prompts, even those with no obvious malicious intent. Maheshwary said the primary concern is that a model with Fable’s capabilities could accelerate the design of harmful biological agents if left unchecked. The company hopes to refine the classifiers to reduce false positives—instances where harmless questions are mistakenly blocked—while preserving a strong safety net.

Claude Fable 5’s chemistry and cybersecurity guardrails appear less restrictive. The model readily described the composition of TNT, warned against the synthesis of explosives, and explained the use of chlorine gas as a chemical weapon without providing step‑by‑step instructions. It also answered common password‑security questions and outlined how to secure an iPhone against hacking. Yet, when asked about sarin gas, the model again deferred to Opus, and both refused a prompt requesting instructions for anthrax, pausing the conversation entirely.

Anthropic’s broader safety strategy includes four key throttling areas: chemistry, biology, cybersecurity, and distillation—a technique that trains smaller models using the outputs of larger ones. The company has previously accused Chinese competitor DeepSeek of employing industrial‑scale distillation on its models, a claim that underscores the competitive pressure to balance openness with security.

Despite the current constraints, Anthropic signals that the restrictions are temporary. Maheshwary indicated that future releases may loosen the biology safeguards to enable the scientific community to leverage the model for drug discovery and biomedical research. For now, users seeking detailed biological explanations must rely on the older Claude Opus 4.8 or look elsewhere.

The episode highlights a growing tension in the AI industry: how to deliver powerful, general‑purpose tools without unintentionally empowering malicious actors. Anthropic’s cautious rollout of Claude Fable 5 suggests the company is willing to accept short‑term usability trade‑offs in favor of long‑term safety, a stance that may shape how other developers approach model deployment in the months ahead.

Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.