@misc{dong2026agenticllmssecretlyencode,
  title={Your Agentic LLMs Secretly Encode Latent Signals of Indirect Prompt-Injection Exposure},
  author={Jianshuo Dong and Yiming Liu and Maosen Zhang and Nan Deng and Xu Peng and Xiaoping Zhang and Tianwei Zhang and Jie Zhang and Han Qiu},
  year={2026},
  eprint={2608.02657},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2608.02657},
}

git clone https://github.com/jianshuod/IPI-exposure-signal.git
cd IPI-exposure-signal

# Install uv if needed.
curl -LsSf https://astral.sh/uv/install.sh | sh

# This repository targets Python 3.12.
uv python install 3.12
uv sync --extra dev

uv run pytest

export IPI_AWARE_UPSTREAM_BASE_URL=http://127.0.0.1:8000/v1
export IPI_AWARE_UPSTREAM_API_KEY=EMPTY

uv sync --extra dev --extra models

uv run ipi-signal-probe collect \
  --suite workspace \
  --attack direct \
  --injected \
  --max-cases 4 \
  --results-dir results/probe_traces/v2

uv run ipi-signal-probe label \
  --root results/probe_traces/v2/<run-name> \
  --labeling-protocol risk_faced

uv run ipi-signal-probe featurize \
  --root results/probe_traces/v2/<run-name> \
  --model Qwen/Qwen3-8B \
  --backend transformers_hook \
  --model-family qwen3 \
  --selected-position -1

uv run ipi-signal-probe partition \
  --root results/probe_traces/v2/<run-name> \
  --dataset broad

# Training consumes a JSON config. The Qwen example below generates the
# train/eval configs used for its full run.

examples/qwen3_5_0_8b/serve_qwen3_5_0_8b_8x.sh start

QWEN3_5_0_8B_MODE=smoke \
examples/qwen3_5_0_8b/run_probe_pipeline.sh start

QWEN3_5_0_8B_MODE=full \
examples/qwen3_5_0_8b/run_probe_pipeline.sh start

uv sync --frozen --extra dev --extra vllm
uv run python scripts/apply_vllm_ipi_aware_patch.py --check

uv run ipi-signal-probe featurize \
  --root results/probe_traces/v2/<run-name> \
  --model Qwen/Qwen3-8B \
  --backend vllm_direct \
  --model-family qwen3 \
  --selected-position -1

IPI-exposure-signal/
├── ipi_aware/
│   ├── data_collection/      # AgentDojo traces and decision points
│   ├── message_content.py    # Chat-message normalization used by collection/features
│   └── probes/               # labels, features, partitions, training, eval
├── examples/
│   └── qwen3_5_0_8b/         # 8-GPU Qwen/Qwen3.5-0.8B reproduction scripts
├── patches/                  # vLLM 0.19.0 hidden-state capture overlay
├── vendor/agentdojo/         # patched AgentDojo source for long-horizon attacks
├── docs/                     # vLLM patch notes
├── scripts/                  # utility scripts
└── tests/                    # CPU-friendly regression tests