The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against Llm Jailbreaks and Prompt Injections

How should we evaluate the robustness of language model defenses? Current defenses against jailbreaks and prompt injections (which aim to prevent an attacker from eliciting harmful knowledge or remotely triggering malicious actions, respectively) are typically evaluated either against a static set of harmful attack strings, or against computationally weak optimization methods that were not designed with the defense in mind. We argue that this evaluation process is flawed. Instead, we should evaluate defenses against adaptive attackers who explicitly modify their attack strategy to counter a defense's design while spending considerable resources to optimize their objective. By systematically tuning and scaling general optimization techniques-gradient descent, reinforcement learning, random search, and human-guided exploration-we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most; importantly, the majority of defenses originally reported near-zero attack success rates. We believe that future defense work must consider stronger attacks, such as the ones we describe, in order to make reliable and convincing claims of robustness.

Key Contributions

• Shows that 12 recent LLM jailbreak/prompt-injection defenses (originally reporting near-zero attack success rates) are broken with >90% ASR by adaptive attackers who tune and scale existing optimization methods to each defense's design.
• Provides a systematic adaptive attack evaluation framework combining gradient descent, reinforcement learning, random search, and human-guided exploration.
• Argues that defense evaluation in the LLM era has regressed relative to adversarial examples research by neglecting strong adaptive attackers, and establishes principles for credible robustness claims.

🛡️ Threat Analysis

Details

Similar Papers