Outputs are the content a model produces after it is prompted or deployed: text, code, images, audio, or structured data. In security work, outputs matter because they are the model’s visible behavior and the point where errors, policy violations, or malicious use become operational risk.
Attackers may try to shape outputs into phishing messages, harmful code, leaked secrets, or policy-bypassing instructions. Defenders review outputs for hallucinations, data exposure, toxic content, and signs of prompt injection or jailbreaks. Output filtering, human review, logging, and provenance checks help reduce misuse and show whether a system is generating safe, authorized material.