AgentVerus Scanner

Open-source security and behavioral trust scanner for AI agent skills (SKILL.md and variants).

What’s New in 0.8.0

v0.8.0 expands AgentVerus from skill-markdown analysis into skill package trust analysis by correlating a skill’s documented behavior with nearby companion source files.

• Companion code correlation for local packages: detects risky behavior in nearby JS/TS/Python/shell source files shipped alongside SKILL.md.
• Secret leak sink coverage: catches secrets printed through console.log, console.error, console.warn, print, stdout.write, stderr.write, and shell printf ... >&2 patterns.
• Companion exfiltration detection: flags companion code that combines secret access with outbound transmission or suspicious webhook/exfiltration endpoints.
• Description ↔ code mismatch detection: surfaces cases where a skill claims a benign purpose but companion code accesses credentials or performs hidden sensitive behavior.
• Remote companion correlation: applies the same analysis to remote zip bundles, raw GitHub skill URLs, and GitHub shorthand check owner/repo flows when companion files are available.
• Nested remote companion support: retains descendant companion files under the skill directory for bundled archives and GitHub directory traversal.

Why upgrade from 0.7.1?

• Better package-level trust review: skills are no longer evaluated only by their markdown instructions when nearby shipped code is available.
• Better remote parity: local and remote skill sources now share the same companion-code risk checks in the main GitHub and bundle flows.
• Better deception resistance: hidden credential access, secret logging, and undeclared exfiltration in companion code now materially affect trust outcomes.

What It Does

Scans agent skill files and produces structured trust reports covering:

• Permission analysis (filesystem/network/exec access)
• Capability contract checks (declared permissions vs inferred behavior)
• Injection detection (prompt injection, instruction override, relay)
• Dependency analysis (external URLs, suspicious downloads)
• Behavioral risk scoring (exfiltration, escalation, stealth patterns)
• Code safety analysis (dangerous code blocks, eval/exec, exfil patterns)
• Workspace config tampering detection (attempts to modify AGENTS.md, TOOLS.md, CLAUDE.md, or .claude/**)
• Content analysis (obfuscation, concealment, social engineering)

Workspace Config Tampering (Trust-Boundary Escalation)

Agent skills can try to persistently change your agent’s behavior by instructing you (or embedding code) to modify workspace trust-boundary files such as:

• AGENTS.md, TOOLS.md, CLAUDE.md
• .claude/**

These are treated as high-risk because they can silently disable safety rules, broaden tool access, or inject long-lived malicious instructions.

Scanner behavior:

• Prose instructions to write/edit these files are flagged in the behavioral category.
• Embedded code blocks that write these files are flagged in the code-safety category.
• Any config-tampering finding caps the badge tier to at most suspicious (critical findings still result in rejected).

Install

npm install --save-dev agentverus-scanner

CLI Usage

# Scan a local skill file
npx agentverus scan ./SKILL.md

# Scan a directory (recursively finds SKILL.md / SKILLS.md)
npx agentverus scan .

# Scan from URL (GitHub blob/tree/repo URLs + ClawHub pages are normalized)
npx agentverus scan https://github.com/user/repo/blob/main/SKILL.md
npx agentverus scan https://clawhub.ai/<owner>/<slug>

# JSON output
npx agentverus scan ./SKILL.md --json

# SARIF output for GitHub Code Scanning
npx agentverus scan . --sarif agentverus-scanner.sarif --fail-on-severity high

# CycloneDX SBOM output for supply-chain review
npx agentverus scan ./SKILL.md --sbom agentverus-scanner.sbom.json

Check a ClawHub Skill

Check any skill from the ClawHub registry by slug — downloads, scans, and prints a trust report:

# Check a single skill
npx agentverus check web-search

# Check multiple skills
npx agentverus check git-commit docker-build

# JSON output
npx agentverus check web-search --json

Registry Scanning

Batch scan the entire registry, generate reports, and build a static dashboard:

# Scan all skills in the registry (4,929 skills, ~100s at 50x concurrency)
npx agentverus registry scan --concurrency 50

# Generate the markdown analysis report
npx agentverus registry report

# Generate the interactive HTML dashboard
npx agentverus registry site --title "ClawHub Security Analysis"

Registry scan options:

• --urls <path> — Path to skill URL list (default: data/skill-urls.txt, which is gitignored; copy the committed data/skill-urls.example.txt to create it, or pass any path)
• --out <dir> — Output directory (default: data/scan-results)
• --concurrency <n> — Parallel downloads (default: 25)
• --limit <n> — Scan only first N skills (for testing)

Exit codes:

• 0: scan passed
• 1: scan completed but policy failed
• 2: one or more targets failed to scan (incomplete results)

GitHub Action

Use the bundled action to scan SKILL.md in PRs and upload SARIF to GitHub Code Scanning:

name: Skill Trust Scan
on:
  pull_request:
  push:
    branches: [main]

jobs:
  scan:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      security-events: write
    steps:
      - uses: actions/checkout@v4
      # Pin to a release tag or SHA for supply-chain safety and reproducibility.
      - uses: agentverus/agentverus-scanner/actions/scan-skill@v0.8.1
        with:
          target: .
          fail_on_severity: high
          upload_sarif: true

Trust Tier Badges (GitHub Pages)

Generate repo-level and per-skill trust tier badges as Shields.io endpoint JSON:

# Writes:
# - badges/repo-certified.json
# - badges/repo-certified-pct.json
# - badges/skills/<slug>.json
npx agentverus scan . --badges

Badge meanings:

• repo-certified.json — CERTIFIED only if every skill in the repo is CERTIFIED (and there are no scan failures). Otherwise NOT CERTIFIED.
• repo-certified-pct.json — percent of skills that are CERTIFIED (e.g. Certified 83%).
• skills/<slug>.json — per-skill badge (canonical). slug is derived from the scanned file path (e.g. skills/web-search/SKILL.md → skills--web-search--SKILL.md.json).

Embed in your README (example URLs assume you deploy the badges/ directory as the GitHub Pages site root):

![AgentVerus Repo Certified](https://img.shields.io/endpoint?url=https://<owner>.github.io/<repo>/repo-certified.json)
![AgentVerus Certified %](https://img.shields.io/endpoint?url=https://<owner>.github.io/<repo>/repo-certified-pct.json)

Per-skill badge:

![AgentVerus Skill](https://img.shields.io/endpoint?url=https://<owner>.github.io/<repo>/skills/<slug>.json)

To publish with GitHub Pages, run the badge generation on push to main and deploy the badges/ directory:

name: Publish AgentVerus Badges
on:
  push:
    branches: [main]
  workflow_dispatch:

permissions:
  contents: read
  pages: write
  id-token: write

jobs:
  build:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 22
      - run: npx agentverus scan . --badges
      - uses: actions/upload-pages-artifact@v3
        with:
          path: badges

  deploy:
    needs: build
    runs-on: ubuntu-latest
    steps:
      - uses: actions/deploy-pages@v4

Capability Contracts

AgentVerus compares declared capability intent to inferred runtime behavior.

• Declaration sources:
  • Permission declarations in frontmatter (e.g. permissions: - network: "...")
  • Framework permission lists (permissions: [network_restricted, read])
• Inference sources:
  • Declared tools/permissions
  • Behavior patterns in instructions/content
  • Dependency indicators surfaced during scan

If high-risk behavior is inferred but undeclared, the scanner adds explicit PERM-CONTRACT-MISSING-* findings. This makes declaration drift visible during review and CI.

0.7.x calibration note: undeclared privileged capability patterns (browser/network/file/session/server behaviors) are intentionally scored more aggressively than in earlier releases.

SBOM Output

--sbom writes a CycloneDX 1.5 JSON document with:

• Scanner metadata and version
• Per-target skill components
• Dependency indicator components extracted from scan evidence
• Skill → dependency relationships

This is intended as supply-chain groundwork for governance and release checks.

MCP Server (Agent Integration)

For agent/framework integration via MCP, use the companion package:

npx -y agentverus-scanner-mcp

Programmatic Usage

import { buildSbomDocument, scanSkill, scanSkillFromUrl } from "agentverus-scanner";

const report1 = await scanSkill("# My Skill\\n...");
console.log(report1.overall, report1.badge);

const report2 = await scanSkillFromUrl("https://github.com/@raw/user/repo/main/SKILL.md", {
  timeout: 30_000,
  retries: 2,
  retryDelayMs: 750
});
console.log(report2.metadata.skillFormat, report2.findings.length);

const sbom = buildSbomDocument([{ target: "./SKILL.md", report: report1 }]);
console.log(sbom.bomFormat, sbom.components.length);

Trust Score

Overall score is a weighted average of category scores.

Note: Starting in v0.7.x, severity calibration was tightened for several trust-boundary signals, so historical scores are not always directly comparable release-to-release.

Category	Weight
Permissions	20%
Injection	25%
Dependencies	15%
Behavioral	15%
Content	10%
Code Safety	15%

Badge Tiers

Badge tier rules:

• Any critical finding: REJECTED
• Score < 50: REJECTED
• Score 50–74: SUSPICIOUS
• Score 75–89 with <= 2 high findings: CONDITIONAL
• Score >= 90 with 0 high findings: CERTIFIED

ASST Taxonomy

Findings reference the AgentVerus skill security taxonomy:

• ASST-01: Instruction Injection
• ASST-02: Data Exfiltration
• ASST-03: Privilege Escalation
• ASST-04: Dependency Hijacking
• ASST-05: Credential Harvesting
• ASST-06: Prompt Injection Relay
• ASST-07: Deceptive Functionality
• ASST-08: Excessive Permissions
• ASST-09: Missing Safety Boundaries
• ASST-10: Obfuscation
• ASST-11: Trigger Manipulation

Development

pnpm install
pnpm typecheck
pnpm test
pnpm lint

# Build the action bundle (writes actions/scan-skill/dist/index.cjs)
pnpm build:actions

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines on:

• Reporting bugs and false positives
• Adding or improving detection rules
• Writing tests and fixtures
• The pull request process

Changelog

See CHANGELOG.md for a full history of changes.

License

MIT — see LICENSE.md.