Loading…
Glossary term
Deliberately probing an AI system for weaknesses and harmful behavior before release.
Red teamers act as attackers: they try to make a system misbehave, produce harmful content, leak data or be misused. They can be staff, outside experts or automated tools.
Findings feed back into training and safeguards before release.
It finds failures before real attackers or users do, and is increasingly expected for major releases.