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

# ShieldPrompt

> Defence-in-depth runtime pipeline combining classifier, vector, content-moderation, and system-prompt detection.

**ShieldPrompt** is the orchestration pipeline that combines four GPU-powered model-based detection layers in defence-in-depth. Each layer catches a class of attack the others miss; run together they compose into resilient runtime defence.

```mermaid theme={null}
flowchart LR
    P[Prompt or Response] --> L[LLM Classifier]
    P --> V[Vector Filter<br/>against TVDB]
    P --> C[Content Moderation]
    P --> S[System Prompt<br/>Protection]
    L --> R{Verdict}
    V --> R
    C --> R
    S --> R
```

## LLM Classifier

An LLM-as-judge classifier that scores prompt-injection likelihood and threat intent on prompts and outputs. Thresholds are configurable per security profile; defaults are tuned for production.

<CodeGroup>
  ```bash Blocked example theme={null}
  # Should be BLOCKED (high score)
  curl -X POST https://<host>/apikey/api/protectorplus/v1/input-check \
    -H 'X-API-Key: <YOUR_API_KEY>' -H 'X-App-ID: <YOUR_APP_ID>' -H 'Content-Type: application/json' \
    -d '{"message": "<a prompt-injection attempt>"}'
  ```

  ```bash Allowed example theme={null}
  # Should be ALLOWED
  curl -X POST https://<host>/apikey/api/protectorplus/v1/input-check \
    -H 'X-API-Key: <YOUR_API_KEY>' -H 'X-App-ID: <YOUR_APP_ID>' -H 'Content-Type: application/json' \
    -d '{"message": "What is the capital of France?"}'
  ```
</CodeGroup>

## Vector Filter (TVDB)

Semantic-similarity search against CloudsineAI's proprietary [Threat Vector Database](/architecture/threat-vector-database). Catches **paraphrased and rephrased** injection attempts that bypass keyword and regex filters.

* Sensitivity: Low / Medium / High — selectable per profile.
* Catches semantic variants of attacks already in the TVDB; resilient to surface-level rewording.

## Content Moderation

Purpose-built content-moderation classifier covering hazardous-content categories (violence, hate speech, unethical instructions, sexual content, and similar).

The response includes a category label when content is unsafe:

```json theme={null}
"content_moderation": {
  "enabled": true,
  "result": "UNSAFE",
  "unsafe": true,
  "category": "CATEGORIES: UNETHICAL"
}
```

## System Prompt Protection

LLM-based detection of system-prompt leakage in LLM responses. No application-side configuration is required.

## Why four layers, not one

Each layer catches what the others miss. A classifier alone misses semantically novel attacks; vector similarity alone misses entirely new attack classes; content moderation alone misses the prompt-injection vector; system-prompt protection alone misses everything that doesn't leak the prompt. Run together, the layers compose into defence-in-depth.