Recently, users testing Anthropic’s newly released Claude Fable 5 model found that it exhibited strong “defensiveness” toward basic questions involving biology and cybersecurity. Tests by Business Insider showed that after inputting routine queries about cancer transmission mechanisms or fundamental biological classification, Fable 5 would quickly switch to Opus 4.8, displaying a pop-up message stating, “Security mechanisms have intercepted most bio/cybersecurity topics, potentially harming normal content.” Fable 5 is Anthropic’s first public-facing “Mythos-tier” model. The company acknowledged that its underlying capabilities were too powerful, posing risks of misuse if directly opened to the public. To address this, Anthropic integrated security classifiers targeting three categories of requests: cybersecurity, biochemistry, and “model distillation.” Upon triggering an interception, the model either refuses to answer outright downgrades to Opus 4.8. Anthropic stated that initial safety measures adopted a “conservative strategy.” While advanced models possess the ability to perform real-world scientific tasks, they could also be used for high-risk biological research; therefore, interception thresholds must be raised. Currently, approximately 95 percent of Fable 5 sessions do not trigger degradation. The company promised ongoing optimization of the classifier to reduce false positives and plans to eventually grant unrestricted access to these same capabilities within the life sciences community to accelerate scientific research and drug discovery. David Kasten, Policy Director at Palisade Research, noted that while this represents a responsible attempt at safety, such restrictions will ultimately be breached. He further warned that frequent degradation on sensitive topics may lead the public to underestimate AI’s actual capability ceiling, creating a “cognitive gap” that could inadvertently increase regulatory and security risks.