A blocking classifier is an automated policy filter used to decide whether an AI system should answer a request, refuse it, or route it for review. It typically evaluates the prompt, conversation context, and sometimes the model’s planned output against safety rules such as malware creation, credential theft, phishing, or exploit guidance.
In cyber security, blocking classifiers help create a trust boundary around high-risk AI tools. They reduce accidental leakage of offensive instructions, but they can also create false positives that interrupt legitimate defensive work. That is why some approved programs use reduced safeguards for vetted users, while still keeping logging and oversight in place. Attackers may try to evade these filters with obfuscation, encoding, roleplay, or prompt injection. Defenders respond by tightening policy logic, adding abuse monitoring, and reviewing blocked requests to improve the classifier over time.