A serious incident is a harmful event linked to an AI system that goes beyond a minor error. It can include injury, major disruption, discrimination, privacy or rights harm, or other regulated damage. In AI security and compliance, the key question is not only whether the model was hacked, but whether its output or action plausibly contributed to real-world harm.
This matters because incident response must track causality, preserve logs, model versions, prompts, outputs, and human approvals, and decide quickly whether reporting duties apply. In practice, a complaint, audit finding, or business anomaly may be the first signal. Defenders use triage, rollback, and evidence preservation to confirm scope and limit impact. In attacks, adversaries may exploit model errors, poisoned inputs, or workflow integration to trigger harmful outcomes without a traditional breach.