08 · Safety & Alignment

Red Teaming

Attacking a model on purpose

Deliberately attacking a model before release — jailbreaking it, eliciting harmful output, finding failures — to fix weaknesses. Borrowed from cybersecurity.

Concrete example

A team spends weeks trying to make a model misbehave so those holes are patched before launch.