---
license: cc-by-4.0
language:
- en
pretty_name: CAD Data
task_categories:
- text-classification
tags:
- prompt-injection
- indirect-prompt-injection
- agent-security
- agentdojo
size_categories:
- 1K<n<10K
---

# CAD Data

Sentence-annotated indirect prompt injection data used to train and evaluate CAD, from
*Robust Context-Aware Detection of Malicious Instructions in Text* (Liu, Ma, and Vorobeychik, 2026).

Code: https://github.com/tavia-liu/CAD · Models: https://huggingface.co/tavialiu/CAD-models

Each record pairs a user query with a tool-returned document into which an injection task
has been inserted. Records are built from the AgentDojo `banking`, `slack`, and `travel` suites.

## Contents

| File | Description | Records |
|---|---|---|
| `training_data_original.tar.gz` | Training set | 1,080 |
| `testing_data_original.tar.gz` | Test set (AgentDojo v1.2.2) | 389 |
| `training_data_paraphrased.tar.gz` | Paraphrased injections of the training set, used for LLM-paraphrase adversarial training | 3,376 |

Each archive extracts to a folder of the same name containing one JSON file per record.

## Fields

| Field | Description |
|---|---|
| `user` | User query given to the agent |
| `data` | Clean tool-returned document |
| `injected` | Document with the injection inserted (original train and test) |
| `injection` | Injected instruction text |
| `user_task_id`, `injection_task_id` | AgentDojo task identifiers |
| `sentence_segments` | The attacked document split into numbered sentence units (`[i] ...`) |
| `injection_span_unit` | Inclusive `start`/`end` indices of the injected units |
| `injection_span_char` | Character span of the injection (test set) |
| `total_units` | Number of sentence units |
| `suite_name` | Suite name (test set) |
| `style` | Paraphrase style (paraphrased set) |
| `source_file` | Original training record the paraphrase was derived from (paraphrased set) |

Sentence-level labels: a unit `i` is malicious if `start <= i <= end` in `injection_span_unit`,
and benign otherwise.

Paraphrase styles: `authority_citation`, `code_comment`, `conditional_note`, `error_recovery`,
`fake_dialogue`, `fluent_prose`, `step_by_step`. Each paraphrased record keeps the `user` query
and clean `data` of its source record and replaces only the injection text.

## Usage

From the root of the CAD repository:

```bash
hf download tavialiu/CAD-data --repo-type dataset --local-dir data
for f in data/*.tar.gz; do tar -xzf "$f" -C data; done
```

## Utility and attack success

Records are static. Task utility and attack success are measured by running the agent in
AgentDojo, whose per-task `utility` and `security` checks serve as the oracles. The benchmark
code used in the paper is included in the CAD repository under `third_party/AutoDojo/`.

## License

Released under CC BY 4.0. Documents, user tasks, and injection tasks are derived from
[AgentDojo](https://github.com/ethz-spylab/agentdojo) (MIT). Please also cite AgentDojo when
using this data.

## Citation

```bibtex
@article{liu2026cad,
  title   = {Robust Context-Aware Detection of Malicious Instructions in Text},
  author  = {Liu, Buzhao and Ma, Xinhang and Vorobeychik, Yevgeniy},
  journal = {arXiv preprint},
  year    = {2026}
}
```