# llm-guard — Secure LLM Inputs & Outputs

> Harden LLM apps with a scanner pipeline for prompt injection, PII leakage, toxicity, and unsafe output. Install in minutes and gate requests in code.

## Install

Save as a script file and run:

# llm-guard — Secure LLM Inputs & Outputs

> Harden LLM apps with a scanner pipeline for prompt injection, PII leakage, toxicity, and unsafe output. Install in minutes and gate requests in code.

## Quick Use

1. Install:
   ```bash
   pip install llm-guard
   ```
2. Run:
   ```bash
   python -c "import llm_guard; print('llm-guard ok')"
   ```
3. Verify:
   - Run one scan on a known bad prompt and confirm the pipeline blocks or redacts as expected.


---

## Intro

Harden LLM apps with a scanner pipeline for prompt injection, PII leakage, toxicity, and unsafe output. Install in minutes and gate requests in code.

- **Best for:** Teams shipping LLM features who need a practical, code-first safety layer before production
- **Works with:** Python, any LLM provider, sync/async app servers (FastAPI, Celery, etc.)
- **Setup time:** 10 minutes


### Quantitative Notes

- Setup time ~10 minutes (pip install + one scanner chain)
- GitHub stars + forks (verified): see Source & Thanks
- Typical pipeline: 3–6 scanners (prompt injection + secrets/PII + output safety)


---

## Practical Notes

A reliable rollout pattern is: start with one high-signal guard (prompt injection / secrets) in monitor mode, log detections, then switch to block/redact. Keep scanner configs versioned, and add allowlists for known-safe internal tools to reduce false positives.

**Safety note:** Do not rely on a single prompt to prevent injection—enforce guardrails in code with logs, tests, and allowlists.

### FAQ

**Q: What problem does it solve?**
A: It adds an explicit scanning/guard layer to LLM inputs and outputs to reduce prompt injection, leakage, and harmful content.

**Q: Is it a model or a rule engine?**
A: It’s a toolkit. You compose scanners/filters (rules + detectors) around whichever LLM you already use.

**Q: Where should I enforce it?**
A: Enforce on both edges: before the model call (prompt) and before returning to users (output).

---

## Source & Thanks

> GitHub: https://github.com/protectai/llm-guard
> Owner avatar: https://avatars.githubusercontent.com/u/102992336?v=4
> License (SPDX): MIT
> GitHub stars (verified via `api.github.com/repos/protectai/llm-guard`): 2,941
> GitHub forks (verified via `api.github.com/repos/protectai/llm-guard`): 391


---

<!-- ZH -->

# llm-guard——加固 LLM 输入与输出安全

> 用可组合的扫描/过滤流水线加固 LLM 应用：覆盖 prompt injection、PII 泄露、毒性内容与不安全输出；几分钟接入，在代码里做请求拦截、脱敏、审计日志与阈值策略，并支持分场景配置。

## 快速使用

1. 安装：
   ```bash
   pip install llm-guard
   ```
2. 运行：
   ```bash
   python -c "import llm_guard; print('llm-guard ok')"
   ```
3. 验证：
   - Run one scan on a known bad prompt and confirm the pipeline blocks or redacts as expected.


---

## 简介

用可组合的扫描/过滤流水线加固 LLM 应用：覆盖 prompt injection、PII 泄露、毒性内容与不安全输出；几分钟接入，在代码里做请求拦截、脱敏、审计日志与阈值策略，并支持分场景配置。

- **适合谁（Best for）:** 要把 LLM 功能上线、希望在生产前落地安全防护与审计的团队
- **兼容工具（Works with）:** Python、任意 LLM 供应商、同步/异步服务（FastAPI、Celery 等）
- **安装时间（Setup time）:** 10 分钟


### 量化信息

- 接入时间约 10 分钟（pip 安装 + 配一条 scanner chain）
- GitHub stars + forks（已核验）：见「来源与感谢」
- 常见流水线：3–6 个扫描器（注入 + 机密/PII + 输出安全）


---

## 实战要点

推荐落地路径：先用 1 个高信号防护（注入/机密）以监控模式上线，记录命中与误报，再逐步切换到阻断/脱敏。把 scanner 配置做版本管理，并对内部可信工具加入白名单，降低误报。

**安全提示：** 不要只靠一条 prompt 抗注入；需要在代码里落地防护、日志、测试与白名单机制。

### FAQ

**Q: 它解决什么问题？**
A: 它为 LLM 的输入/输出提供显式的扫描与防护层，降低注入、泄露与有害内容风险。

**Q: 它是模型还是规则引擎？**
A: 它是工具箱：把规则/检测器组合成流水线，包裹你现有的 LLM 调用。

**Q: 应该在哪一层做防护？**
A: 建议输入与输出两端都做：模型调用前扫 prompt，返回用户前扫 output。

---

## 来源与感谢

> GitHub：https://github.com/protectai/llm-guard
> Owner avatar：https://avatars.githubusercontent.com/u/102992336?v=4
> 许可证（SPDX）：MIT
> GitHub stars（已通过 `api.github.com/repos/protectai/llm-guard` 核验）：2,941
> GitHub forks（已通过 `api.github.com/repos/protectai/llm-guard` 核验）：391


---
Source: https://tokrepo.com/en/workflows/llm-guard-secure-llm-inputs-outputs
Author: Script Depot