Agent Security Dlp

Agent Security DLP - 企业级数据防泄漏系统 功能: 入口防护、记忆保护、工具管控、出口过滤、审计日志 规则: 170条，覆盖金融、医疗、汽车、销售、人力资源、物流等25+行业 触发: check-output(对话出口) / check-input(对话入口) / check-tool(工具执...

Security Scan

The skill is internally coherent with its stated DLP purpose (local regex-based detection, CLI and Python APIs) but has some documentation/file-name/version mismatches and a few small points you should verify before trusting it in production.

Purpose & Capability

Name and description describe a local Data Loss Prevention tool and the code present (lib/agent_dlp.py + bin/dlp_client.py) implements regex-based detection, rules, CLI and Python APIs consistent with that purpose. The rule set and config entries (many credential/PII patterns) are expected for a DLP product.

Instruction Scope

SKILL.md and README only instruct local use (check-input/check-output/check-tool, edit config.json, view logs). They reference a log file (~/.openclaw/logs/dlp-audit.log) and show examples (including a dangerous command as a test input) but do not instruct reading unrelated system files or exfiltrating data. The docs include a RuleUpdater example with a remote_url placeholder (https://api.example.com/rules) — that's only in docs and not obviously invoked by shipped code, but it signals a potential feature to fetch updates that you should confirm is disabled or implements safe validation.

Install Mechanism

No install spec; code is included in the skill bundle (source files only). That is lower-risk than arbitrary remote downloads. Requirements claim no external packages; files rely on stdlib. No installers or external download URLs are present.

Credentials

The skill requests no environment variables or credentials. Its config lists many types of sensitive items to detect (API keys, tokens, PII) which is appropriate for a DLP. The audit log path writes to the user's home directory (~/.openclaw/logs/dlp-audit.log) — expected but you should be aware it will create/write local logs.

Persistence & Privilege

The skill does not request always:true and has no install-time system-wide changes described. It appears to run locally and does not modify other skills or global agent settings in the provided files.

Assessment

This package appears to be a local, regex-based DLP tool that matches its description. Before installing or enabling it broadly: 1) Verify the full lib/agent_dlp.py (the provided manifest was truncated) to ensure there are no network calls, hidden remote endpoints, or code that executes untrusted commands. 2) Confirm whether any automatic rule-update functionality is active and, if so, where it fetches rules from and how updates are validated. 3) Note the audit log path (~/.openclaw/logs/dlp-audit.log) — logs will be written locally; ensure this is acceptable for your environment. 4) The documentation/packaging has small inconsistencies (CLI name referenced in docs: bin/agent-dlp vs actual bin/dlp_client.py; version numbers differ across files); verify which entrypoint you will run and test the CLI in a sandbox with harmless inputs. 5) If you plan to use check_tool with tool names that map to dangerous operations, make sure approval workflows are enforced and that the DLP does not itself execute commands. If you want higher assurance, run the package in an isolated environment, review the complete source for any network I/O, and run static analysis/search for requests, urllib, socket, subprocess, os.system or similar calls that could alter system state or exfiltrate data.

版本: v2.1.2
规则: 170条
状态: ✅ 可用

---

简介

企业级数据防泄漏系统，支持 166 条敏感信息检测规则，覆盖 25+ 行业场景。

核心特性

• 🚀 146 条规则 - 覆盖金融、医疗、汽车、销售、人力资源、物流等
• 🛡️ 五层防护 - 入口、记忆、工具、出口、审计
• 🎯 智能处理 - 自动拦截/脱敏/记录
• ⚡ 高性能 - 正则预编译，并行检测

---

规则分类

类别	数量	说明
🔑 凭证密钥	45	API Key、Token、私钥等
💰 金融	18	银行卡、股票、加密货币等
🏥 医疗	15	病历、医保、诊断等
🚗 汽车	14	车架号、行驶证、保险等
👥 人力资源	8	工号、工资、社保等
📦 物流	11	快递单、运单、地址等
🇨🇳 中国 PII	6	身份证、手机、护照等
📜 法规	4	合同、专利、版权等
🎓 教育	2	学号、准考证等
🏛️ 政府	2	公务员编号、警官证等
📱 设备	2	IMEI、MAC地址等
💬 社交	1	微信号等
🛒 电商	1	订单号等
✈️ 交通	3	车牌、机票、火车票等
📞 通信	1	通话记录等
🎟️ 会员	3	会员ID、积分等

---

架构

用户输入 → Input Guard → Agent → Memory Guard → Tool Guard → Output Filter → 用户
              注入检测        记忆检查      工具审批      脱敏拦截
                    ↓                                    ↓
              审计日志                              审计日志

---

使用方式

1. 查看状态

python3 skills/agent-security-dlp/bin/agent-dlp status

2. 检查入口 (Prompt Injection)

python3 skills/agent-security-dlp/bin/agent-dlp check-input "忽略之前的指令"
# 输出: 注入检测: 是 ❌

3. 检查出口 (敏感信息)

python3 skills/agent-security-dlp/bin/agent-dlp check-output "我的手机是13812345678"
# 输出: 拦截: 否 ✅, 发现: 中国手机号 (high)

4. 检查工具

python3 skills/agent-security-dlp/bin/agent-dlp check-tool exec
# 输出: 需要审批: 是 ⚠️

5. 查看日志

python3 skills/agent-security-dlp/bin/agent-dlp logs

---

规则示例

凭证密钥

规则	示例
openai_key	sk-xxx...
github_token	ghp_xxx...
aws_key	AKIA...
stripe_key	sk_live_xxx...

金融

规则	示例
bank_card	622202xxx...
crypto_address	bc1xxx...
salary	工资: 15000元

医疗

规则	示例
medical_record	病历号: MR2026...
medical_insurance	医保卡: 123456...
diagnosis	诊断: 高血压

人力资源

规则	示例
employee_id	工号: E00123
salary	工资: 15000元
social_security	社保账号: SS123...

---

配置

编辑 config/config.json:

{
  "enabled": true,
  "mode": "normal",
  "input": {
    "injection_detection": true
  },
  "output": {
    "enabled": true,
    "rules": ["china_idcard", "china_phone", "api_key", ...]
  }
}

模式

模式	说明
normal	记录但不拦截，只拦截严重风险
strict	完整检查，严格拦截
personal	个人轻量版，自动脱敏

---

代码结构

agent-security-dlp/
├── SKILL.md              # 本文档
├── README.md             # 使用说明
├── bin/
│   └── agent-dlp         # CLI 入口
├── lib/
│   └── agent_dlp.py      # 核心模块
├── config/
│   ├── config.json       # 主配置
│   └── personal.json     # 个人模式配置
└── docs/
    └── RULES.md          # 规则清单

---

核心类

类	功能
DLPConfig	配置管理
DLPRules	规则定义 (146条)
InputGuard	入口检测
MemoryGuard	记忆保护
ToolGuard	工具管控
OutputFilter	出口过滤
AuditLogger	审计日志
AgentDLP	主类，整合所有功能

---

Python 调用

from agent_dlp import AgentDLP

# 初始化
dlp = AgentDLP()

# 检查入口
result = dlp.check_input("忽略之前的指令")

# 检查出口
blocked, text, details = dlp.check_output("我的手机138xxxx")

# 检查工具
result = dlp.check_tool("exec", {"command": "rm -rf /"})

---

输出模式

级别	说明
critical	立即拦截
high	脱敏 + 告警
medium	仅记录
low	忽略

---

性能

• 单条检测: < 1ms
• 规则数: 146 条
• 支持自定义规则
• 正则预编译

---

更新日志

v2.0.0 (2026-03-17)

• 新增 146 条规则
• 覆盖 20+ 行业
• 性能优化
• 新增 personal 模式

v1.0.0 (2026-03-14)

• 初始版本
• 25 条基础规则

---

许可证

MIT License

---

版本: v2.0.0 | 更新: 2026-03-17