Prompt filtering is a defensive control that inspects prompts, tool instructions, or generated output for suspicious, unsafe, or policy-violating content. In AI systems, it is used to block obvious abuse such as jailbreak attempts, requests for secrets, or commands that try to steer an agent into risky actions.
It matters because agents can turn text into execution. If an attacker slips malicious instructions into a prompt, document, email, or web page, the agent may follow them unless the system filters the input first. Filtering can reduce exposure and catch low-effort attacks, but it is not authorization. A strong defense pairs prompt filtering with least privilege, identity-based access control, and human approval for high-impact actions, so a dangerous prompt cannot become a dangerous action.