Deliberately attempting to break or mislead an AI model to identify vulnerabilities and failure modes.