Why Anthropic's 'safe' Mythos-class model won't answer questions about cancer

Anthropic has introduced its latest AI model, Claude Fable 5, which boasts significant capabilities but comes with stringent safety measures. The company has acknowledged that these safeguards, designed to prevent misuse, may hinder the model's utility in certain fields like cybersecurity and biology. The release of Claude Fable 5 follows the company’s earlier decision to restrict access to its original Mythos model due to serious cybersecurity concerns. According to Anthropic, the powerful nature of the Mythos-class models necessitated the implementation of broad safety protocols to ensure responsible use by the public. Users have reported that when querying Fable 5 on topics such as cancer or online misinformation, the model frequently switches to a less sophisticated version, Opus 4.8, citing safety measures. This is due to the model's built-in classifiers that flag a range of queries related to biology and cybersecurity, which can sometimes lead to the blocking of benign content. In a statement, a spokesperson from Anthropic explained that these precautionary measures are crucial for the safe deployment of AI technology in sensitive areas where malicious use could lead to significant harm. The company has categorized flagged requests into three main areas: cybersecurity, biology and chemistry, and the capabilities of Fable 5 itself. Despite these limitations, Anthropic asserts that the safety measures are a necessary step in the responsible development of AI. Early data suggests that over 95% of sessions with Fable 5 do not trigger a switch to the less capable model, although the company admits that false positives can occur. Anthropic aims to refine these safeguards over time, with the goal of eventually allowing the unrestricted use of Mythos-class models in biological research and drug development. The conversation surrounding AI safety continues to evolve, with experts like David Kasten from Palisade Research highlighting the ongoing tension between innovation and security. He noted that while the safeguards represent a proactive approach, they also reflect a broader concern about the risks associated with increasingly advanced AI technologies. The challenge remains to balance the powerful capabilities of AI with the imperative to prevent potential misuse.