A research paper proposes an auditing framework to stress-test prompt injection defenses for large language models before release, revealing weaknesses in current defenses and highlighting the importance of pre-deployment evaluation.