Prompt-injection defenses are usually evaluated as binary detectors, but deployed LLM applications must make operational decisions about whether to allow, block, or escalate potentially malicious inputs. A new paper on arXiv argues that this mismatch limits the practical value of current detection systems.

The study proposes a framework that moves from detection to decision, incorporating calibration and deferral. Calibration ensures that the detector's confidence scores reflect true likelihoods, while deferral allows the system to escalate uncertain cases rather than forcing a binary choice.

This approach acknowledges that in real deployments, the cost of a wrong decision varies. By integrating probabilistic reasoning and a deferral option, the framework aims to make prompt-injection defenses more robust and more aligned with how LLM applications actually operate.