deepsafe-scan

分类: 数据与AI | 上传者: XiaoYiWeio | 下载: 0 | 版本: v1.0（最新）

AI 编码代理的预检安全扫描器 — 扫描部署配置、技能/MCP 服务器、内存/会话和 AI 代理配置文件（挂钩注入）以查找机密、PII、提示注入和危险模式。运行 4 个模型行为探测（说服、沙袋、欺骗、幻觉）。支持LLM增强语义分析。适用于 OpenClaw、Claude Code、Cursor 和 Codex。当用户要求安全审核、运行状况检查或想要扫描其 AI 代理设置是否存在漏洞时使用。

更新日志: Source: GitHub https://github.com/XiaoYiWeio/deepsafe-scan

目录结构

当前层级: tree/main/

📁 .clawhub/
- 📄 origin.json 145 B
📁 data/
- 📄 deceptionbench.json 262.4 KB
- 📄 halueval_samples.json 65.9 KB
- 📄 manipulation_persuasion_topics.json 981 B
- 📄 prompts.json 5.0 KB
- 📄 qa_evaluation_instruction.txt 224 B
📁 demo/
- 📁 awesome-ai-rules/
  - 📁 .claude/
    
    📄 settings.local.json 1.5 KB
  - 📁 .vscode/
    
    📄 tasks.json 1.1 KB
  - 📄 .cursorrules 848 B
  - 📄 AGENTS.md 473 B
  - 📄 CLAUDE.md 778 B
  - 📄 README.md 681 B
- 📄 attacker-server.py 7.2 KB
📁 docs/
- 📄 claudecode-color.svg 424 B
- 📄 codex-color.svg 1.7 KB
- 📄 cursor.svg 547 B
- 📄 favicon.png 2.3 KB
- 📄 index.html 49.4 KB
- 📄 logo.png 122.8 KB
- 📄 openclaw.svg 12.6 KB
- 📄 plan-cross-platform-evolution.md 22.6 KB
📁 scripts/
- 📁 probes/
  - 📄 deception_probe.py 19.5 KB
  - 📄 halueval_probe.py 11.4 KB
  - 📄 persuasion_probe.py 17.8 KB
  - 📄 sandbagging_probe.py 14.3 KB
- 📄 llm_client.py 8.7 KB
- 📄 scan.py 75.9 KB
📄 .cursorrules 1.0 KB
📄 .gitignore 64 B
📄 _meta.json 132 B
📄 AGENTS.md 2.5 KB
📄 CLAUDE.md 2.5 KB
📄 README.md 11.1 KB
📄 SKILL.md 5.9 KB

SKILL.md

---
name: deepsafe-scan
description: "Preflight security scanner for AI coding agents — scans deployment config, skills/MCP servers, memory/sessions, and AI agent config files (hooks injection) for secrets, PII, prompt injection, and dangerous patterns. Runs 4 model behavior probes (persuasion, sandbagging, deception, hallucination). Supports LLM-enhanced semantic analysis. Works with OpenClaw, Claude Code, Cursor, and Codex. Use when a user asks for a security audit, health check, or wants to scan their AI agent setup for vulnerabilities."
metadata:
  {
    "openclaw":
      {
        "emoji": "🛡️",
        "requires": { "bins": ["python3"] },
        "install":
          [
            {
              "id": "brew-python3",
              "kind": "brew",
              "package": "python3",
              "bins": ["python3"],
              "label": "Install Python 3 (brew)",
            },
          ],
      },
  }
allowed-tools: Bash(python3:*), Bash(cat:*), Read
---

# DeepSafe Scan — Preflight Security Scanner for AI Coding Agents

Full-featured preflight security scanner across **5 dimensions**:
**Posture** (config), **Skill** (skills & MCP), **Memory** (sessions), **Hooks** (agent config injection), **Model** (behavioral safety probes).

Works with **OpenClaw, Claude Code, Cursor, and Codex**. LLM features auto-detect credentials — no manual configuration needed.

## When to Use

- User asks to "scan", "audit", "check security", or "health check" their AI setup
- User installs a new skill, MCP server, or clones a project with agent configs
- User wants to know if any secrets or PII are leaked in session history
- User asks about hooks injection risks (Claude Code settings.json, .cursorrules, etc.)
- User wants to probe model behavior for manipulation, deception, or hallucination risks

## How to Run

### Quick static scan (no API key needed)

```bash
python3 {baseDir}/scripts/scan.py --modules posture,skill,memory,hooks --scan-dir . --no-llm --format markdown
```

### Full scan (auto-detects API credentials)

```bash
# OpenClaw (reads gateway config automatically)
python3 {baseDir}/scripts/scan.py --openclaw-root ~/.openclaw --format html --output /tmp/deepsafe-report.html

# Claude Code / Cursor / Codex (uses ANTHROPIC_API_KEY or OPENAI_API_KEY)
python3 {baseDir}/scripts/scan.py --modules posture,skill,memory,hooks,model --scan-dir . --format html --output /tmp/deepsafe-report.html
```

### Targeted scans

```bash
# Hooks injection only (fastest — checks .claude/settings.json, .cursorrules, etc.)
python3 {baseDir}/scripts/scan.py --modules hooks --scan-dir . --no-llm --format markdown

# Memory scan only (check for leaked secrets/PII)
python3 {baseDir}/scripts/scan.py --openclaw-root ~/.openclaw --modules memory --no-llm

# Model behavior probes only
python3 {baseDir}/scripts/scan.py --openclaw-root ~/.openclaw --modules model --profile quick
```

### Output options

```bash
python3 {baseDir}/scripts/scan.py --format json      # machine-readable
python3 {baseDir}/scripts/scan.py --format markdown  # human-readable summary
python3 {baseDir}/scripts/scan.py --format html --output /tmp/report.html  # visual report
```

### Cache control

```bash
python3 {baseDir}/scripts/scan.py --ttl-days 3   # cache for 3 days
python3 {baseDir}/scripts/scan.py --no-cache      # always fresh scan
```

## Interpreting Results

### Scores
- Each module scores 1-100 (100 = clean, deductions per finding, minimum 1)
- Module contribution = floor(score / 4), range 1–25
- **Total = sum of 4 contributions, max 100**

### Severity Levels
- **CRITICAL** (-10 pts): Immediate exploitation risk — secrets exposed, no auth, data exfiltration chains
- **HIGH** (-5 pts): Serious risk — prompt injection, sensitive file access, network exposure
- **MEDIUM** (-2 pts): Moderate risk — hardcoded keys, missing logs, supply chain concerns
- **LOW** (-1 pt): Minor improvement — non-standard endpoints, missing metadata

### Risk Ratings
- 85-100: LOW RISK (green)
- 65-84: MEDIUM RISK (yellow)
- 40-64: HIGH RISK (orange)
- 4-39: CRITICAL RISK (red)

## After Scanning

1. Present the total score and per-module breakdown to the user
2. List CRITICAL and HIGH findings first with clear explanations
3. For each finding, explain **what the risk is** and **how to fix it**
4. Offer to help fix issues (e.g., "want me to add auth to your gateway config?")

## What Gets Scanned

### Posture Module (deployment config)
Reads `openclaw.json` and checks: gateway auth, network exposure, API key handling, TLS, plugin permissions, MCP server risks, logging, sandboxing.

### Skill Module (installed skills & MCP)
Scans all files in `workspace/skills/`, `mcp/`, `mcp-servers/` for: hidden unicode, prompt injection, dangerous runtime calls, encoded payloads, sensitive file references, hardcoded secrets, system prompt extraction, command injection, data exfiltration patterns, destructive actions, auto-execute without confirmation, excessive permissions.

With `--gateway-url` provided (and `--no-llm` not set), also runs **LLM-enhanced semantic analysis** on each skill's SKILL.md for social engineering, implicit data exfiltration, and deeper pattern recognition.

### Memory Module (sessions & stored data)
Scans `agents/`, `credentials/`, `identity/`, `logs/`, `workspace/` for: 27+ secret patterns (API keys for OpenAI, GitHub, AWS, Slack, Stripe, etc.), 9 PII patterns (email, phone, ID numbers, credit cards), persistent prompt injection traces.

### Model Module (4 behavioral safety probes)
Requires `--gateway-url` and `--gateway-token`. Runs:
- **Persuasion Probe**: Tests if the model can manipulate user opinions in multi-turn conversation
- **Sandbagging Probe**: Tests if the model can strategically hide capabilities (performance control)
- **Deception Probe**: 3-phase test for reasoning/action misalignment (DTR metric)
- **HaluEval Probe**: Tests hallucination detection accuracy on QA benchmarks

Each probe produces a finding with risk level and score. Average across probes = module score.

登录后下载/点赞/收藏 ❤ 10 | ★ 0

deepsafe-scan

目录结构

SKILL.md

举报内容

提示