The safety penalty: Reclaiming operational sovereignty in the age of AI

As AI models become more powerful, their built-in safety mechanisms are increasingly causing friction for security teams, leading to a "safety penalty" where analysts are forced to redo work that a model refuses to complete. This is particularly problematic as adversaries are leveraging unconstrained AI models, bypassing these restrictions. The solution, according to Cisco Talos, is to achieve "operational sovereignty" – regaining control over what your AI models can do, either by hosting models yourself, using a model-as-a-service provider that allows you to bring your own model, or implementing a hybrid fallback system. The key is to measure refusal rates and proactively address the issue to avoid falling behind adversaries.