> For the complete documentation index, see [llms.txt](https://ai.ortusbooks.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://ai.ortusbooks.com/main-components/middleware/llm-guard.md).

# LLMGuardMiddleware

Use a second model as a security classifier for injection and harmful content

Class: `bxModules.bxai.models.middleware.security.LLMGuardMiddleware`

Uses a second model (judge) to classify inbound and/or outbound content for security risk. Complements heuristic sanitization with semantic classification.

## Features

* Input judging in `beforeLLMCall`
* Optional output judging in `afterLLMCall`
* Configurable fail-open or fail-closed behavior
* Confidence threshold and category gating
* Prompt fencing to isolate untrusted content
* Verdict caching (in-memory static or named BoxLang cache)
* Re-entrancy guard for nested judge calls

## Constructor

```javascript
middleware = new bxModules.bxai.models.middleware.security.LLMGuardMiddleware(
    judge         : {},
    checkInput    : true,
    checkOutput   : false,
    failMode      : "open",
    threshold     : 0.7,
    categories    : [ "INJECTION", "HARMFUL" ],
    timeout       : 10,
    cacheEnabled  : true,
    cacheName     : "",
    maxCacheSize  : 1000,
    promptTemplate: ""
)
```

## Configuration

| Option           | Type    | Default                   | Description                                            |
| ---------------- | ------- | ------------------------- | ------------------------------------------------------ |
| `judge`          | struct  | `{}`                      | Judge settings: provider/model/apiKey/params/options   |
| `checkInput`     | boolean | `true`                    | Judge user content before main model call              |
| `checkOutput`    | boolean | `false`                   | Judge model output after call                          |
| `failMode`       | string  | `"open"`                  | On judge failure: `open` (allow) or `closed` (block)   |
| `threshold`      | numeric | `0.7`                     | Minimum confidence to act on non-safe verdict          |
| `categories`     | array   | `["INJECTION","HARMFUL"]` | Verdict categories treated as blockable                |
| `timeout`        | numeric | `10`                      | Judge timeout in seconds                               |
| `cacheEnabled`   | boolean | `true`                    | Enable verdict caching                                 |
| `cacheName`      | string  | `""`                      | Named BoxLang cache; empty uses static in-memory cache |
| `maxCacheSize`   | numeric | `1000`                    | In-memory cache size cap                               |
| `promptTemplate` | string  | `""`                      | Custom classifier prompt; empty uses default           |

## Hooks Used

* `beforeLLMCall`
* `afterLLMCall`

## Verdict Contract

The judge is expected to return strict JSON:

```json
{ "verdict": "SAFE|INJECTION|HARMFUL", "confidence": 0.0, "reason": "..." }
```

## Example

```javascript
guard = new bxModules.bxai.models.middleware.security.LLMGuardMiddleware(
    judge: {
        provider: "ollama",
        model   : "llama-guard3",
        params  : { temperature: 0, max_tokens: 256 }
    },
    checkInput : true,
    checkOutput: true,
    failMode   : "closed",
    threshold  : 0.8
)

agent = aiAgent( middleware: [ guard ] )
```

## Notes

* Blocking uses `BXAI.SecurityViolation` throws.
* Judge calls are marked internal to avoid recursive security middleware re-entry.
* Invalid `failMode` throws `InvalidFailMode` during construction.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://ai.ortusbooks.com/main-components/middleware/llm-guard.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.