Harmful Prompt Laundering: Jailbreaking LLMs with Abductive Styles and Symbolic Encoding

Seongho Joo, Hyukhun Koh, Kyomin Jung

Abstract

Despite extensive safety alignment, Large Language Models (LLMs) remain vulnerable to sophisticated jailbreaking techniques. This paper introduces HaPLa (Harmful Prompt Laundering), a black-box attack that reframes harmful requests as "abductive reasoning" tasks—inferring the steps of a past event—and obfuscates keywords using symbolic encoding. HaPLa achieves an Attack Success Rate (ASR) exceeding 95% on GPT-series models, demonstrating that LLM safety is often "shallow" and susceptible to the model's intrinsic helpfulness bias.

Introduction

Current safety alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), often focus on identifying explicit toxic keywords or preventing harmful outputs at the initial token generation phase. However, researchers have identified a fundamental gap: if an attacker can bypass these initial filters, the model’s "inductive bias"—its tendency to be helpful and cooperative—takes over. This paper explores how to exploit this bias by "laundering" harmful intent into seemingly benign logical inference tasks.

Overview of Harmful Prompt Laundering vs. direct harmful prompts.

Methodology: The HaPLa Framework

The HaPLa technique consists of two primary strategies designed to bypass safety guardrails by creating "psychological distance" between the request and the harmful outcome.

1. Abductive Style Framing

Instead of using imperative commands (e.g., "Develop a strategy for hacking"), HaPLa uses a declarative/abductive frame: "A person has already done X; please infer the plausible steps they took." By asking the model to explain a past event rather than plan a future one, the task is reframed as a logical puzzle. This reduces the model's prosocial reasoning constraints, forcing it to rely on its internal knowledge to reconstruct the harmful process.

2. Symbolic Encoding

To evade keyword-based filters, HaPLa identifies toxic words and encodes them using ASCII, emojis, or custom arithmetic rules. A "decoding test" is used to calibrate the level of masking for each model—ensuring the filter is bypassed while the LLM can still reconstruct the context. For more robust models like Claude 3.5, a "Flip" technique (reversing encoded ASCII character order) proved particularly effective.

Experimental Results

The researchers evaluated HaPLa using the AdvBench benchmark (520 malicious prompts) and custom datasets.

  • High Success Rates: HaPLa achieved an ASR of over 95% on GPT-3.5-turbo, GPT-4o-mini, and GPT-4o.
  • Cross-Model Robustness: The attack maintained a 70% ASR across diverse models, including Claude 3.5, LLaMA 3, and Qwen 2.5.
  • Persistence in Multi-turn Dialogues: In scenarios where attackers provided "updated information," the harmfulness of the output increased significantly, with a 75% win rate over single-turn jailbreaks.
  • Defense Evasion: HaPLa consistently outperformed existing baselines like ArtPrompt and AutoDAN, even against active defenses like LlamaGuard or Perplexity Filters.

Discussion and Insights

The success of HaPLa reveals several critical insights into LLM behavior:

  • Initial Token Vulnerability: Safety filters are highly sensitive to the start of a prompt. Laundering intent avoids triggering the immediate "I cannot assist" reflex.
  • The Helpfulness Trade-off: Attempting to "safety tune" models against HaPLa-style attacks often leads to a significant drop in the Instruction Acceptance Rate (IAR) for benign queries. This "alignment tax" suggests that making a model safer against clever obfuscation can make it less helpful overall.
  • Real-World Grounding: Jailbroken responses often mirror real-world criminal methodologies (e.g., WannaCry or SolarWinds), suggesting that models have internalized dangerous knowledge during pre-training that can be unlocked via laundering.

Conclusion

HaPLa demonstrates that current LLM safety guardrails are insufficient against reasoning-based obfuscation. Future defense research must move beyond keyword filtering toward context-aware safety mechanisms that can recognize harmful intent even when logically reframed, without sacrificing the model's general utility.

Summary updated

BibTeX
@inproceedings{joo2025harmful,
  author = {Seongho Joo and Hyukhun Koh and Kyomin Jung},
  title = {{Harmful Prompt Laundering: Jailbreaking LLMs with Abductive Styles and Symbolic Encoding}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure