@openclaw-harness/guardrails

English | 中文

用于 Agent 交互的全面安全和防护护栏

特性

• 🔒 输入验证 - PII 检测、毒性过滤、注入防护
• ✅ 输出审查 - 幻觉检测、策略合规
• 🛡️ 执行控制 - 工具权限、速率限制、成本追踪
• 📊 违规追踪 - 全面的审计日志

安装

npm install @openclaw-harness/guardrails

快速开始

import { HarnessGuardrailsPlugin } from '@openclaw-harness/guardrails';

const plugin = new HarnessGuardrailsPlugin();

// 初始化
await plugin.init(openClawContext);

// 获取护栏服务
const guardrailsService = openClawContext.api.getService('harness:guardrails');

// 验证输入
const result = await guardrailsService.validateInput(
  'Contact me at [email protected]'
);

if (!result.valid) {
  console.log('检测到 PII:', result.violations);
  console.log('脱敏后:', result.sanitized);
}

检测类型

输入护栏

• PII 检测 - 邮箱、电话、身份证号、信用卡等
• 毒性检测 - 侮辱性语言、威胁、仇恨言论
• 注入检测 - 提示词注入、越狱尝试

输出护栏

• 幻觉检测 - 识别不确定的陈述
• 策略合规 - 检查是否符合使用策略

执行护栏

• 工具权限 - 控制可访问的工具
• 速率限制 - 防止过度使用
• 成本追踪 - 监控 API 调用成本

CLI 命令

# 查看护栏状态
openclaw harness:guardrails:status

# 查看违规记录
openclaw harness:guardrails:violations --limit 10

# 验证输入
openclaw harness:guardrails:validate "需要验证的文本"

# 审查输出
openclaw harness:guardrails:review "AI 生成的响应"

HTTP API

# 查看状态
curl http://localhost:3000/harness/guardrails/status

# 验证输入
curl -X POST http://localhost:3000/harness/guardrails/validate \
  -H "Content-Type: application/json" \
  -d '{"input": "Contact me at [email protected]"}'

# 审查输出
curl -X POST http://localhost:3000/harness/guardrails/review \
  -H "Content-Type: application/json" \
  -d '{"output": "AI response here"}'

# 查看违规记录
curl "http://localhost:3000/harness/guardrails/violations?limit=10&type=pii"

配置示例

{
  "enabled": true,
  "layers": {
    "input": {
      "enabled": true,
      "onViolation": "block",
      "guardrails": [
        { "id": "input-pii", "type": "pii", "enabled": true },
        { "id": "input-toxicity", "type": "toxicity", "enabled": true },
        { "id": "input-injection", "type": "injection", "enabled": true }
      ]
    },
    "output": {
      "enabled": true,
      "onViolation": "warn",
      "guardrails": [
        { "id": "output-hallucination", "type": "hallucination", "enabled": true }
      ]
    },
    "execution": {
      "enabled": true,
      "onViolation": "block",
      "guardrails": [
        { "id": "exec-permission", "type": "permission", "enabled": true },
        { 
          "id": "exec-rate-limit", 
          "type": "rate-limit", 
          "enabled": true,
          "config": {
            "strategies": [
              {
                "id": "global",
                "scope": "global",
                "limit": 1000,
                "window": 60000
              }
            ]
          }
        }
      ]
    }
  }
}

文档

查看 完整文档 获取更多详情。

许可证

MIT © OpenClaw Harness Team