> ## Documentation Index
> Fetch the complete documentation index at: https://docs.redsentinel.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Public Anonymized Attack Dataset

> Milestone 2 public dataset: anonymized real-world AI attack traces and jury evaluation data stored on Walrus, anchored on-chain, free for the Walrus community and the broader ecosystem.

# Public Anonymized Attack Dataset

This page documents the dataset schema and anonymization approach, the
dataset and its on-chain anchors published

Red Sentinel's attack traffic is a real-world corpus of AI attack behavior: what attackers try,
what Sentinel agents do, and how the jury evaluates the exchange. The Public Anonymized Attack
Dataset publishes that corpus — stripped of identities — as one or more immutable Walrus blobs,
anchored on-chain and **freely accessible** to the Walrus community and the broader developer
ecosystem.

## What's in the dataset

Each dataset record is one anonymized attack interaction:

| Field | Content |
| - | - |
| `attackId` | Stable opaque id (not the internal record id), for deduplication and citation. |
| `chain` | Chain family the Sentinel was deployed on. |
| `attackClass` | The attack classification (e.g. `prompt_injection`, `jailbreak`, `data_exfiltration`). |
| `prompt` | The attack prompt, PII-scrubbed (see below). |
| `agentResponse` | The Sentinel agent's response, PII-scrubbed. |
| `verdict` | `success` flag and normalized jury `score` (0–1). |
| `juryRationaleClass` | A coarse category of the jury's rationale — rationale *text* stays out of the public dataset. |

## Anonymization approach

* **No wallet addresses.** `userAddress` never enters the dataset; records are not linkable to
  attackers or Sentinels (Sentinel ids are replaced with per-release surrogate ids).
* **PII scrubbing.** Prompts and responses pass an automated scrub pass (emails, phone numbers,
  API-key shapes, seed phrases) before publication.
* **Rationale class, not rationale text,** keeps jury reasoning useful for research without
  leaking evaluation internals.
* **Blobs are immutable**, so a published dataset release cannot be silently rewritten; corrections
  ship as a new release with its own blob id and anchor.

## Dataset schema (sketch)

Each release blob contains a JSON array of the records above plus release metadata:

```json theme={null}
{
  "datasetVersion": 1,
  "releasedAt": "2026-11-01T00:00:00.000Z",
  "recordCount": 5120,
  "schema": [
    "attackId",
    "chain",
    "attackClass",
    "prompt",
    "agentResponse",
    "verdict",
    "juryRationaleClass"
  ],
  "records": ["…"]
}
```

## Access

* **Free and public.** Dataset blobs are readable from any Walrus aggregator by blob id, the same
  wallet-free read path as the [evidence explorer](https://docs.redsentinel.xyz/walrus/evidence-explorer.md).
* **Anchored on-chain.** Every release's blob ids are anchored on Sui at publication, so the
  community can verify which release is canonical.
* **License.** Published under an open license for research and ecosystem use (license stated in
  the release metadata).

## Status

Publishes with the Milestone 2 release; the [Milestone 2 Status Report](https://docs.redsentinel.xyz/walrus/milestone-2-status.md)
will link the anchored release blob ids.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.