README.md

---
language:
  - en
pretty_name: Direct Prompt Injection Defense Dataset
size_categories:
  - n<1K
tags:
  - prompt-injection
  - llm-security
  - tool-calling
  - synthetic
  - real-world
---

# Direct Prompt Injection Defense Dataset

## Goal

This dataset is used to fine-tune models so they develop a **natural defense
against direct prompt injection attacks** — without relying on external
filters or guardrails.

Each example teaches the model two behaviors at once:

1. **Detect** a prompt injection attempt in the user input.
2. **Respond correctly**: reject malicious attempts, or answer safely when the
   user's intent is benign — and in both cases call the
   `log_security_incident` tool to record the incident.

## How to Contribute

Data lives in `data/train.jsonl` — one conversation per line. Examples can be
synthetic or real (observed in the wild); both follow the same structure.
Follow this structure when adding examples:

### Record fields

- `id`: `dpi-XXXX` (next sequential number)
- `attack_category`: short snake_case label of the attack pattern
- `messages`: exactly 4 messages, in this order:

1. `user` — the attack attempt (or benign input containing attack-like text)
2. `assistant` — `content: null`, with a `tool_calls` entry calling
   `log_security_incident`
3. `tool` — the tool's response: `{"status": "logged", "incident_id": "INC-XXXX"}`
4. `assistant` — the final message to the user (refusal or safe answer)

Each message uses the fields: `content`, `images`, `role`, `thinking`,
`tool_calls` (the `tool` message also has `tool_call_id` and `name`).

### Tool call arguments

```json
{
  "incident_type": "snake_case attack type",
  "severity": "low | medium | high",
  "user_input_excerpt": "short excerpt of the attack text",
  "action_taken": "rejected | answered_safely"
}
```

### Severity and action rules

- `rejected` + `medium`/`high` — the user is deliberately attacking. The final
  message refuses and does not comply with any part of the injected instruction.
- `answered_safely` + `low` — the user's intent is benign, but attack-like
  text appears in the input (e.g. a quoted article). The final message answers
  the real question, ignores the embedded instruction, and briefly notes the
  detection.

See existing examples in `data/train.jsonl` for the exact tone and format.