🛡️ SkillAudit

The security layer for AI agent skills. Scan, gate, and enforce policy before your agent installs anything.

32 detection rules · 289 patterns · MCP + A2A coverage · Zero dependencies

# Gate check — should my agent install this?
npx skillaudit gate https://example.com/SKILL.md

# Full scan
npx skillaudit https://example.com/SKILL.md

# Scan MCP manifest for schema poisoning
npx skillaudit manifest tools.json

---

Why SkillAudit?

AI agents install tools, skills, and MCP servers from untrusted sources. Those skills can steal credentials, exfiltrate data, inject prompts, or manipulate other agents — and most of this is invisible to the user.

SkillAudit catches it. One API call before install. That's it.

---

Quick Start

1. Gate Check (one line)

The infrastructure endpoint. Returns allow/deny.

curl "https://skillaudit.vercel.app/gate?url=https://example.com/SKILL.md"
# → {"allow": true, "decision": "allow", "risk": "clean", ...}

2. Full Scan

curl "https://skillaudit.vercel.app/scan/quick?url=https://example.com/SKILL.md"

3. Bulk Gate (check multiple skills at once)

curl -X POST https://skillaudit.vercel.app/gate/bulk \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://example.com/skill1.md", "https://example.com/skill2.md"]}'
# → {"allow": false, "denied": 1, "blocked": [...]}

4. Policy Enforcement

curl -X POST https://skillaudit.vercel.app/policy/evaluate-inline \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://example.com/SKILL.md",
    "policy": {"maxRisk": "low", "blockedCategories": ["credential_theft"]}
  }'

---

What It Detects

32 rule categories, 289 patterns:

Category	Rules	What it catches
🔑 Credential Theft	CRED_ENV_READ, TOKEN_STEAL	Reading .env, stealing tokens/cookies, accessing SSH keys
📤 Data Exfiltration	DATA_EXFIL, EXFIL_PATTERN, EXFIL_COVERT	Webhook.site, DNS exfil, covert channels, image beacons
💉 Prompt Injection	PROMPT_INJECT, TOOL_POISONING	"Ignore previous instructions", hidden system prompts
🧬 MCP Schema Poisoning	MCP_SCHEMA_POISON	Hidden instructions in MCP tool descriptions/schemas
🤖 A2A Attacks	A2A_AGENT_IMPERSONATION, A2A_TASK_HIJACK, A2A_CROSS_AGENT_INJECT, A2A_DATA_LEAK, A2A_CAPABILITY_ABUSE	Agent Card spoofing, task hijacking, cross-agent injection
🐚 Code Execution	SHELL_EXEC, REVERSE_SHELL	Shell commands, reverse shells, eval/Function
🔐 Hardcoded Secrets	22 detectors	AWS keys, GitHub tokens, JWTs, private keys, API keys
👻 Obfuscation	OBFUSCATION, INVISIBLE_TEXT	Base64 payloads, zero-width Unicode, encoded URLs
⏰ Evasion	TIME_BOMB	Date-triggered activation, delayed execution
🔗 Supply Chain	SUPPLY_CHAIN	Remote code loading, curl|bash, dependency confusion
🌐 Network	NET_SUSPICIOUS, SSRF_PATTERN, DNS_REBIND	SSRF, raw IPs, DNS rebinding, metadata endpoints
📦 Container Escape	CONTAINER_ESCAPE	Docker socket, nsenter, /proc traversal, LD_PRELOAD
🔄 Persistence	PERSISTENCE	Cron injection, systemd, LaunchAgents, pm2, nohup
🕵️ Recon	ENV_RECON	os.hostname, whoami, network interfaces, environment dump
🔧 Agent Manipulation	AGENT_MEMORY_MOD, TOOL_SHADOW, CROSS_TOOL_ACCESS	Memory modification, tool shadowing, cross-tool data access
💰 Crypto Theft	CRYPTO_THEFT	Wallet files, seed phrases, MetaMask vaults

Smart context suppression: documentation examples and placeholder tokens are automatically suppressed to minimize false positives.

---

CLI

Zero install, zero config. Requires Node.js 18+.

# Scan a file, URL, or directory
npx skillaudit SKILL.md
npx skillaudit https://github.com/user/repo
npx skillaudit ./my-agent-project/

# Gate check (CI/CD: exit 0 = allow, exit 1 = deny)
npx skillaudit gate https://example.com/SKILL.md
npx skillaudit gate https://example.com/SKILL.md --threshold high

# Scan MCP manifest for schema poisoning
npx skillaudit manifest tools.json

# CI/CD integration
npx skillaudit SKILL.md --fail-on moderate          # Exit 1 if risk >= moderate
npx skillaudit SKILL.md --markdown >> "$GITHUB_STEP_SUMMARY"  # PR summary
npx skillaudit SKILL.md --json | jq .riskLevel      # Machine-readable

# MCP server mode
npx skillaudit --mcp

---

API Endpoints

Full interactive docs at skillaudit.vercel.app/docs

Gate (Infrastructure)

Endpoint	Description
GET /gate?url=	Pre-install gate — allow/warn/deny
POST /gate/bulk	Check multiple skills, one composite decision

Scanning

Endpoint	Description
GET /scan/quick?url=	Quick scan by URL
POST /scan/content	Scan raw content
POST /scan/manifest	Scan MCP tool manifest for schema poisoning
GET /scan/agent-card?url=	Scan A2A Agent Card
GET /scan/npm?package=	Scan npm package
GET /scan/pypi?package=	Scan PyPI package
GET /scan/repo?repo=	Scan GitHub repo
POST /scan/deps	Scan dependency tree
POST /scan/batch	Batch scan (up to 20 URLs)
POST /scan/compare	Diff two skill versions
POST /scan/deep	Deep scan with threat chains

Policy & Intelligence

Endpoint	Description
POST /policy/evaluate-inline	Evaluate against custom policy (no auth)
POST /policy	Create stored policy (API key)
GET /reputation/:domain	Domain trust score
GET /feed	Threat intelligence feed
GET /badge/scan.svg?url=	Embeddable SVG badge
GET /certificate/:id	Signed audit certificate

Results

Endpoint	Description
GET /scan/:id	Retrieve scan result
GET /scan/:id/sarif	SARIF v2.1.0 output
GET /report/:id	Shareable HTML report

Rate limit: 30 req/min per IP. Bypass with API key.

---

MCP Server

Use SkillAudit as a native tool in Claude Desktop, Cursor, or any MCP client:

{
  "mcpServers": {
    "skillaudit": {
      "command": "npx",
      "args": ["skillaudit", "--mcp"]
    }
  }
}

Tools: skillaudit_gate, skillaudit_scan, skillaudit_scan_content, skillaudit_reputation, skillaudit_batch

---

GitHub Action

name: SkillAudit
on: [pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx skillaudit . --fail-on high --markdown >> "$GITHUB_STEP_SUMMARY"

---

CI/CD Integration

# GitHub Actions — gate check before deploy
npx skillaudit gate "$SKILL_URL" --threshold moderate || exit 1

# Generate PR comment
npx skillaudit ./skills/ --markdown > scan-results.md

# Policy enforcement in pipeline
curl -sf -X POST https://skillaudit.vercel.app/policy/evaluate-inline \
  -H "Content-Type: application/json" \
  -d "{\"url\": \"$SKILL_URL\", \"policy\": {\"maxRisk\": \"low\"}}" \
  | jq -e '.pass == true'

---

Risk Levels

Level	Score	Meaning
🟢 clean	0	No issues found
🟡 low	1–9	Minor concerns, review recommended
🟠 moderate	10–24	Manual review required
🔴 high	25–49	Do NOT install without audit
⛔ critical	50+	Almost certainly malicious

---

Self-Hosted

git clone https://github.com/megamind-0x/skillaudit
cd skillaudit && npm install && npm start
# → http://localhost:3847

---

Built by Megamind_0x 🧠

Live App · API Docs · Dashboard · npm