> ## Documentation Index
> Fetch the complete documentation index at: https://neuraltrust-92b43583-develop.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent & MCP security detectors

> Indirect prompt injection detection for MCP tool results and other tool-sourced content.

Agent / MCP security detectors protect tool-using agents from **indirect prompt
injection** in content that comes back from tools (MCP tool results, tool-role
messages), not from the end-user prompt alone.

| Detector                  | Slug                        | Sides         | Protocols | Backend              |
| ------------------------- | --------------------------- | ------------- | --------- | -------------------- |
| Indirect Prompt Injection | `indirect_prompt_injection` | input, output | llm, mcp  | NeuralTrust Firewall |

## Indirect Prompt Injection — `indirect_prompt_injection`

Scores MCP tool results (and other tool-sourced content) with the NeuralTrust
Firewall indirect prompt injection detector. Preferentially inspects
`role=tool` messages; falls back to all messages or the request body.

Reports a finding when the max category score is at or above threshold
(`signal.type` is typically `injection` or the top-scoring category).

| Field                      | Type   | Required | Notes                                                           |
| -------------------------- | ------ | -------- | --------------------------------------------------------------- |
| `injection.threshold`      | number | ✅        | Score in `[0, 1]` above which content is flagged.               |
| `provider` / `credentials` | —      | —        | Firewall provider + credentials (same pattern as Prompt Guard). |

```json theme={null}
{
  "name": "IPI — tool results",
  "plugin_slug": "indirect_prompt_injection",
  "settings": { "injection": { "threshold": 0.7 } }
}
```

## When to use

* MCP / tool-calling agents that ingest untrusted tool output into the model context.
* Pair with [Prompt Guard](/trustguard/detectors/content-security#prompt-guard--prompt_guard)
  for direct jailbreaks on user prompts.