🤖 AI Summary
Traditional defense mechanisms are vulnerable to model-guided automated attacks due to their predictable rejection feedback, enabling high-success-rate jailbreaks and prompt injection threats against AI systems. This work proposes a novel “mislead-after-detection” paradigm that shifts the defensive objective from complete attack blocking to reducing the efficiency of attackers’ strategy optimization by generating safe yet misleading responses to confound adversarial judgment. Leveraging probabilistic modeling, we design a lightweight Contextual Misleading via Probabilistic Estimation (CMPE) mechanism and theoretically prove that it asymptotically bounds attack success rates. Evaluated on standard jailbreaking benchmarks, CMPE reduces the upper bound of attack success rates by up to two orders of magnitude, effectively eliminating nearly all successful end-to-end automated attacks.
📝 Abstract
Agentic AI systems increasingly rely on language-model components to interpret instructions, process external data, invoke tools, and coordinate with other agents. These capabilities make prompt-injection and jailbreak attacks more consequential, especially as attackers adopt model-guided automation to scale probing, prompt refinement, and response evaluation. This work analyzes the resulting attack-defense setting through a probabilistic model of a target system, its defense mechanism, and the attacker's automated judge. Our analysis shows that conventional detect-and-block defenses can allow attacker success rate (ASR) to approach one as the query budget grows, since predictable refusals provide useful feedback to automated search. We then examine detect-and-misdirect, where detected malicious interactions receive controlled, non-operational responses designed to induce false-positive errors in the attacker's judge. This strategy reduces the positive predictive value of attacker-selected candidates and yields a bounded asymptotic ASR. We evaluate a proof-of-concept realization of this strategy through Contextual Misdirection via Progressive Engagement (CMPE), a lightweight conversational misdirection method designed to replace predictable refusal text with safe but strategically misleading responses in automated jailbreak settings. On jailbreak benchmarks, CMPE reduces estimated ASR upper bounds by up to two orders of magnitude and nearly eliminates verified attack success in end-to-end PAIR and GPTFuzz attack runs.