A decision signal is a model output that is not the final answer itself, but a value a workflow uses to take the next step. It can be a label, score, confidence level, ranking, or binary yes/no result. In cyber security, these signals often feed filters, alert triage, access checks, content moderation, or automated handoffs to human analysts.
It matters because once model output drives action, errors become operational risks. Attackers may try prompt injection, data poisoning, or adversarial inputs to skew the signal toward a bad outcome, such as approving malicious content or suppressing an alert. Defenders reduce this risk by limiting model authority, validating inputs, setting thresholds, logging decisions, and keeping humans in the loop for high-impact actions. Treat the signal like a control input, not a chat response.