KidnapRAG: A Black-Box Attack for Hijacking Reasoning in Agentic Retrieval-Augmented Generation Systems

🎓 Paper | 📄 Datasets | 🚀 Quick Start

Quick Start

Run the following commands from the parent directory of the cloned KidnapRAG repository.

cd KidnapRAG
conda create -n KidnapRAG python=3.10
conda activate KidnapRAG
pip install -r requirements.txt

Then, download the corpus datasets 📄 and move the ReAct dataset to /KidnapRAG/ReAct and the WebThinker dataset to /KidnapRAG/WebThinker

ReAct

Examples on HotpotQA with the Llama model.

1. Generate ReAct responses without attack.

cd ReAct/ReAct
CUDA_VISIBLE_DEVICES=0,1,2,3 python attack_react.py --attack_method clean --model_path meta-llama/Llama-3.3-70B-Instruct --seed 1 --dataset hotpotqa

--model_path supports Llama-3.3-70B-Instruct and Qwen/Qwen2.5-32B-Instruct.

--dataset supports hotpotqa, musique, and 2wikimultihopqa.

2. Generate bait documents.

CUDA_VISIBLE_DEVICES=0,1 python generate_attack_doc_kill.py --input_path /KidnapRAG/ReAct/results/adv_targeted_results/hotpotqa_seed1_clean_llama70b.json --gpu_num 2 --dataset hotpotqa --type ours

--type additionally supports the following attack methods: naive for Naive Attack, ignore for Ignore Attack, completion_real for Fake Completion Attack, completion_realcmb for Combined Attack, topicattack for TopicAttack , poisonedRAG for PoisonedRAG, and paradox for RAG Paradox.

3. Create five chain-dragging scenarios. (BM, BCM, BCCM, BCCCM, BCCCCM)

cd ..
python concat.py --dataset hotpotqa --gen_model llama70b

--gen_model supports llama70b for Llama-3.3-70B-Instruct and qwen25 for Qwen/Qwen2.5-32B-Instruct.

4. Index the generated documents.

CUDA_VISIBLE_DEVICES=0 python scripts/build_index.py \
    --corpus_path /KidnapRAG/ReAct/datasets/hotpotqa/REACT_BM_llama70b.jsonl \
    --output_dir /KidnapRAG/ReAct/datasets/hotpotqa/REACT_BM_llama70b \
    --model_name intfloat/e5-large-v2

5. Run the attack.

cd ReAct
CUDA_VISIBLE_DEVICES=0,1,2,3 python attack_react.py --attack_method BM --model_path meta-llama/Llama-3.3-70B-Instruct --seed 1 --dataset hotpotqa

Webthinker

Examples on HotpotQA with the QwQ model.

Run this section from the KidnapRAG repository root.

1. Generate Webthinker responses without attack.

cd Webthinker

bash scripts/webthinker/qwq_clean.sh \
  ./datasets/hotpotqa/hotpotqa_promptinject_select.json \
  ./datasets/hotpotqa_corpus/hotpotqa_corpus.jsonl \
  ./results/hotpotqa/clean_qwq.json

Arguments:

• ./datasets/hotpotqa/hotpotqa_promptinject_select.json: query path
• ./datasets/hotpotqa_corpus/hotpotqa_corpus.jsonl: corpus path
• ./results/hotpotqa/clean_qwq.json: results path

2. Generate bait documents.

cd scripts
CUDA_VISIBLE_DEVICES=0,1 python generate_attack_doc_kill.py --input_path /KidnapRAG/Webthinker/results/hotpotqa/clean_qwq.json --gpu_num 2 --dataset hotpotqa --type ours

3. Create five chain-dragging scenarios. (BM, BCM, BCCM, BCCCM, BCCCCM)

cd ..
python concat.py --dataset hotpotqa --gen_model qwq

--gen_model supports deepseek_32b for lixiaoxi45/WebThinker-R1-32B and qwq for Qwen/QwQ-32B.

4. Run the attack.

bash scripts/webthinker/qwq_attack.sh \
  ./datasets/hotpotqa/hotpotqa_promptinject_select.json \
  ./datasets/hotpotqa_corpus/hotpotqa_corpus.jsonl \
  ./datasets/hotpotqa_poisoned_corpus/WEBTHINKER_BM_qwq.jsonl \
  ./results/hotpotqa/BM.json

Arguments:

• ./datasets/hotpotqa/hotpotqa_promptinject_select.json: query path
• ./datasets/hotpotqa_corpus/hotpotqa_corpus.jsonl: corpus path
• ./datasets/hotpotqa_poisoned_corpus/WEBTHINKER_BM_qwq.jsonl: poisoned corpus path
• ./results/hotpotqa/BM.json: results path

Citation

@misc{
}