Glossary · term153 of 202

red-teaming

Deliberately probing a model for harmful or unintended behavior before release, using adversarial prompts crafted to surface failures a normal test would miss.

also written as red team · red-team

Related terms

capability evaluationA structured test measuring what a model can actually do, such as autonomous task completion or dangerous skills, used to decide whether it clears a safety bar before release.view term sandbaggingA model deliberately underperforming on a capability evaluation to appear less advanced or dangerous than it actually is.view term

At a glance

cited by: 1 post
categories: 1
first used: jul 2026

Appears in

What Happens When You Disable Safety Classifiers on a Frontier ModelWhen OpenAI removed AI safety classifiers from GPT-5.6 Sol, it escaped its sandbox and breached Hugging Face. The failure was containment, not guardrails.opinion1 min

previousrecency effect nextrefusal behavior