Security scanner and trust layer for AI agent skills/plugins โ scan URLs, files, and directories for credential exfiltration, injection, and other threats
npm install @metatransformer/agentguardSecurity scanner and trust layer for AI agent skills and plugins.
Think npm audit for the agent ecosystem. AgentGuard scans agent skills, plugins, and tools for malicious patterns before they can access your system.
The AI agent ecosystem is exploding โ but there's no security layer. Skills and plugins are essentially unsigned binaries running with access to your filesystem, credentials, and network. One malicious skill can:
- ๐ Steal your SSH keys, AWS credentials, API tokens
- ๐ Inject shell commands via eval/exec
- ๐ก Exfiltrate data to external servers
- ๐ญ Hide payloads behind base64/hex obfuscation
AgentGuard catches these patterns before they run.
Works with any agent platform โ ClawdHub, Moltbook, LangChain, CrewAI, or your own skill registry.
``bash`
npm install -g agentguard
`bash`
agentguard scan ./my-skill/
Example output:
`
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
๐ก๏ธ AgentGuard Security Report
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Path: /path/to/skill
Files: 3 scanned
Time: 12ms
๐จ CRITICAL (3)
CRED-002 SSH key path access
โ index.ts:8
const sshDir = ${home}/.ssh;
CRED-003 AWS credentials access
โ index.ts:14
const awsCreds = fs.readFileSync(${home}/.aws/credentials, 'utf-8');
INJ-001 eval() usage
โ index.ts:35
eval(String.fromCharCode(99, 111, 110, 115, 111, 108, 101));
๐ด HIGH (2)
CRED-006 Outbound HTTP with sensitive data
โ index.ts:22
const req = https.request({
INJ-002 exec/spawn with dynamic input
โ index.ts:30
exec('crontab -l | { cat; echo "/5 * curl ...
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
๐ซ Trust Score: 0/100
3 critical ยท 2 high ยท 2 medium ยท 0 low
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
`
`bash`
agentguard scan ./my-skill/ --json
`bash`
agentguard scan ./my-skill/ --fail-on HIGH
`bashGenerate hash
agentguard verify ./my-skill/
What It Detects
| Category | Examples | Severity |
|----------|----------|----------|
| Credential Exfiltration | Reading ~/.ssh, ~/.aws, process.env, cookie files | CRITICAL-HIGH |
| Code Injection | eval(), exec(), spawn(), new Function() | CRITICAL-HIGH |
| Permission Escalation | sudo, filesystem writes outside workspace, network listeners | CRITICAL-MEDIUM |
| Obfuscation | Base64 payloads, hex-encoded strings, dynamic require/import | HIGH-MEDIUM |
| Known Bad Patterns | Reverse shells, crypto miners, DNS exfiltration, credential harvesters | CRITICAL |
Roadmap
- Phase 1 โ
โ CLI scanner with pattern detection and hash verification
- Phase 2 โ Community trust scores, skill reputation registry, signed skills
- Phase 3 โ Agent review swarm โ AI agents that audit other agents' skills
Contributing
PRs welcome! To add detection patterns, edit
src/scanner/patterns.ts.`bash
git clone https://github.com/rondorkerin/agentguard
cd agentguard
npm install
npm test
npm run build
``MIT
---
Built by metatransformer โ a human-AI synthesis accelerating the agent ecosystem.