Anthropic's New Fable 5 AI Model Restricts Sensitive Topics

Anthropic has publicly released Claude Fable 5, its latest "Mythos-class" artificial intelligence model. The company states that Fable 5 surpasses its previous frontier Opus models in overall capabilities.

The launch of Fable 5 incorporates safeguards designed to prevent the model from answering queries on specific topics. These restricted areas include cybersecurity, biology, and chemistry. Anthropic has expressed concerns that discussing these subjects could potentially assist malicious actors.

Fable 5 shares its underlying model with Mythos 5, which is emerging from a month-long "Mythos Preview" period. Mythos 5 is currently available only to a small, trusted group of cyberdefenders through Project Glasswing. Unlike Mythos 5, the publicly accessible Fable 5 redirects queries on sensitive topics to the earlier Claude Opus 4.8 model and notifies the user of this action.

Anthropic acknowledges that these safeguards are tuned to be "stricter than ideal," which may occasionally lead to the refusal of "harmless requests." However, the company states that such false positives occur in less than five percent of all testing sessions. This strict approach is considered necessary to prevent scenarios where the model could provide assistance for causing serious harm that might not be obtainable from other sources. (Source: Ars Technica)