A prompt regression test is a repeatable check that runs the same prompts against a model after an update to see whether its answers, refusals, tool use, and formatting still match expectations. It is the AI equivalent of software regression testing: if a new model version or alias changes behavior, the test helps spot it before production users do.
This matters in cyber security because small behavior shifts can break moderation workflows, SOC assistants, fraud review, or code-analysis tools. A model may become less cautious, more willing to call tools, or easier to steer with adversarial prompts. Defenders use prompt regression suites to compare old and new outputs, measure safety thresholds, and catch drift when vendors move aliases or retire models. Attackers may also probe these changes to find newly exposed weaknesses. In practice, teams keep a library of representative prompts and rerun them whenever a model, system prompt, or routing layer changes.