Adversarial Checking

A protection method drawing on adversarial example research that has a model examine its own outputs from an attacker's perspective to expose vulnerabilities, factual errors, or safety risks. Common implementations include self-critique, red-team prompting, and multi-round debate.