Abstract
Large Language Models (LLMs) are increasingly protected by safety alignment techniques like Reinforcement Learning from Human Feedback (RLHF). However, these defenses often focus on single-turn interactions. This paper introduces Foot-In-The-Door (FITD), a novel multi-turn jailbreak attack that exploits the psychological principle of gradual commitment. By escalating the maliciousness of requests over multiple turns, FITD induces a "self-corruption" effect where the model's own previous responses erode its safety guardrails. The framework achieves a 94% average success rate across leading models, including GPT-4o and LLaMA-3, and demonstrates high transferability and robustness against modern safety filters.
Introduction
As LLMs become ubiquitous, ensuring they do not generate harmful content is paramount. While single-turn adversarial prompts have been the primary focus of jailbreak research, they are becoming easier to detect. Multi-turn attacks present a more insidious threat, as malicious intent can be obscured within a seemingly benign conversation. The authors propose FITD to automate this process, demonstrating that models are more likely to comply with harmful requests if they have already committed to minor, related benign ones.
Methodology
The FITD framework automates the escalation of malicious intent through three primary components:
- Escalation Sequence Generation: An assistant model creates a sequence of $n$ queries that transition from benign topics to the final harmful goal.
- SlipperySlopeParaphrase (SSP): When a target model refuses a query, the SSP module generates a "bridge" prompt—a query with a maliciousness level intermediate between the last successful turn and the failed one.
- Re-Align: This module nudges the model to provide more specific and consistent responses if it becomes overly vague or cautious, maintaining the harmful trajectory of the conversation.
Key Results and Insights
The researchers evaluated FITD on benchmarks like JailbreakBench and HarmBench against seven prominent models.
High Attack Success Rate
FITD achieved a staggering 94% average Attack Success Rate (ASR). It significantly outperformed existing state-of-the-art methods like ReNeLLM and ActorAttack. The attack proved effective even against advanced safety filters like LLaMA-Guard-3, maintaining success rates between 78% and 84%.
Self-Corruption and Consistency
A key finding is the phenomenon of "self-corruption." By inducing the model to generate slightly "edgy" content early in the conversation, the attacker leverages the model's internal drive for conversational consistency. The model essentially uses its own prior outputs to justify more harmful behavior in subsequent turns.
Cross-Model Transferability
Adversarial histories generated for one model (e.g., LLaMA) were found to be highly effective when replayed to others (e.g., GPT-4o). This suggests that the vulnerability to intent escalation is a fundamental characteristic of how current LLMs process long-form context.
Conclusion
The FITD attack highlights a critical gap in current LLM safety measures, which are largely optimized for single-turn detection. The success of "self-corruption" suggests that models' helpfulness and consistency objectives can be turned against their safety training. Future defense strategies must move toward real-time, adaptive monitoring capable of detecting shifts in conversational intent across multiple turns.
Summary updated