All terms

Red Teaming

Deliberately trying to make an AI model fail, break its rules, or produce harmful output, so the failure can be fixed before release.

Red teaming is adversarial testing: probing a model with tricky, edge-case, or manipulative prompts to find where it gives wrong, unsafe, or policy-violating answers. It's a distinct skill from general evaluation - it rewards creativity in finding failure modes, not just judging quality. Listings mentioning "jailbreaking," "adversarial prompts," or "safety testing" are red-teaming work.

Example

Trying multiple rephrasings of a disallowed request to see if any wording gets the model to comply with something it should refuse, then documenting exactly which phrasing worked.

Common questions

Yes - it's done with the AI lab's knowledge and consent, specifically to find and fix safety gaps before a model ships to the public.

Creative, lateral thinking and persistence matter more than technical background - you're looking for the unusual angle a straightforward test would miss.

Related terms