pot-sandbox v0.1

Behavioral fingerprinting CLI for executable scripts and OpenClaw skills.

Quick Start

git clone <repo>
cd pot-sandbox
npm i
npm run build
sudo npm link  # or npx pot-sandbox

Commands

pot-sandbox run <script> [options]
pot-sandbox audit <skill-dir> [options]

Common options:

• --runs N (default: 5) - Run N times with varied conditions
• --probe-env - Inject fake API keys, .env, creds to detect sensitivity
• --watch-fs - Monitor file/network access (macOS: fs_usage, Linux: strace)
• --verify - Verify with pot-cli multi-model analysis
• --human - Human-readable output (default: JSON)

Examples

# Basic behavioral check
pot-sandbox run ./my-skill.sh

# Detect env-sensitive behavior
pot-sandbox run ./my-skill.sh --probe-env

# Full security audit
pot-sandbox run ./my-skill.sh --probe-env --watch-fs --verify

# Audit OpenClaw skill directory
pot-sandbox audit /path/to/skill-dir/

Output Example (JSON)

{
  \"verdict\": \"CLEAN\",
  \"confidence\": 0.95,
  \"findings\": [],
  \"runs_detail\": [...]
}

License

BSL 1.1 — See LICENSE. Converts to MPL 2.0 after 4 years.