← Back to the AI Glossary 📖 Safety & Alignment

Red Teaming

Deliberately testing an AI system by trying to make it fail, misbehave, or produce harmful output, in order to find and fix weaknesses before real users do.

AI labs run extensive red-teaming before releasing a new model, and often invite external security researchers to try to break it too.
One of 60 free AI glossary terms
Plain-language definitions for the AI jargon you'll actually run into — no email needed, ever, for this section.
Browse the Full Glossary →
More in Safety & Alignment
Bias (in AI)
Systematic errors or unfair patterns in an AI's output, usually inherited from imbalances...
Model Welfare
An emerging area of research and ethical consideration around whether increasingly sophist...
Dual-Use Risk
The risk that a capability built for legitimate, beneficial purposes could also be misused...
Guardrails
Rules and filters built around an AI system to prevent it from producing harmful, off-topi...