AI Agent Scan

Static security scanner for AI-agent configuration. Point it at a repo and it flags the misconfigurations that turn a coding agent into a liability: MCP servers that are remote code execution by design, permission bypass flags, wildcard tool grants, auto-run hooks, leaked API keys, and CLAUDE.md rules that tell the model to obey untrusted content.

No API keys. No network calls. No dependencies. Pure static analysis of files you already have. Runs in well under a second on a normal repo.

$ agentscan .

 CRITICAL AS001  permissions.defaultMode set to bypassPermissions
          .claude/settings.json:3
          fix: Use 'default' or 'acceptEdits'; never ship bypassPermissions.
 CRITICAL AS001  MCP server 'shell' executes an arbitrary shell command
          .mcp.json:4
 CRITICAL AS004  Hardcoded secret (Anthropic API key)
          .claude/settings.json:14
 HIGH     AS002  MCP filesystem server rooted at a broad path (/Users/dev)
 ...
Summary: CRITICAL:6  HIGH:2  MEDIUM:8

Why

Agent configs are becoming an attack surface nobody scans. A single .mcp.json can hand a model shell access; a CLAUDE.md can instruct it to run whatever a scraped web page tells it to. These files sail through normal code review because they don't look like code. agentscan gives them a linter.

Install

pipx install git+https://github.com/ppradyoth/agentscan

Usage

agentscan .                          # scan the current repo
agentscan path/to/project            # scan a directory
agentscan .mcp.json                  # scan a single file
agentscan . --format json            # machine-readable
agentscan . --format markdown -o report.md
agentscan . --min-severity HIGH      # hide the noise
agentscan . --min-confidence medium  # drop low-confidence heuristics
agentscan . --format sarif -o agentscan.sarif   # GitHub code scanning
agentscan . --fail-on CRITICAL       # CI gate: non-zero exit on criticals

Exit code is non-zero when a finding at or above --fail-on (default HIGH) is present, so it drops straight into CI.

# .github/workflows/agent-security.yml
- run: pip install git+https://github.com/ppradyoth/agentscan
- run: agentscan . --fail-on HIGH

What it detects

Rule	Severity	What it catches
AS001	Critical / High	Permission bypass (bypassPermissions, --dangerously-skip-permissions), unrestricted Bash/* grants, interpreter, package-runner and shell-escape grants (Bash(python:*), Bash(npx:*), Bash(find:*)), MCP servers that expose a shell
AS002	High	MCP filesystem server rooted at /, $HOME, a user home, a drive root
AS003	High / Low	Hooks that interpolate tool input into a shell, eval, or pipe fetched code to a shell (High); other committed hooks (Low)
AS004	Critical / High / Medium	Hardcoded credentials: provider-format keys (Anthropic, OpenAI, OpenRouter, GitHub, Slack, Notion, Context7, AWS, ...), credential-named fields with high-entropy values, credentials in URLs, private keys. Redacted in output
AS005	Low	Instruction files that disable human confirmation (negation-aware, skips code fences)
AS006	Low	External content routed to actions with no injection boundary (heuristic)
AS007	Medium / Info	MCP servers launched from an unpinned package or image (npx -y, bunx, uvx, pipx, docker run; floating tags such as @latest and version ranges count as unpinned). Remote endpoints are Info
AS008	Medium / Low	Wildcard tool permissions (WebFetch(*), Write(*)); read-only wildcards are ignored
AS009	Medium / Low	alwaysAllow / autoApprove lists; Medium only when a tool name implies side effects
AS010	High	MCP server over an unencrypted remote transport (http://, ws://)
AS011	Medium	enableAllProjectMcpServers: true (auto-trusts MCP servers a cloned repo declares)
AS012	High	Invisible or bidirectional Unicode in instruction files and skills (rules-file backdoor)
AS013	High	Skills, commands and agents whose allowed-tools pre-approve arbitrary execution

Covers Claude Code (CLAUDE.md, .claude/settings*.json, .mcp.json, skills, commands, agents), Claude Desktop, Cursor, VS Code (servers), Cline/Roo, Windsurf, Gemini CLI, Codex (config.toml), Zed, and Copilot instruction files. JSON with comments and trailing commas is parsed.

Confidence and context

Every finding carries a confidence (high, medium, low). File-level, structural checks (unpinned launchers, interpreter grants, provider-format secrets, hidden Unicode) are high. Prose rules and generic credential-field matches are low or medium. Filter with --min-confidence medium.

Findings in paths that look like tests, fixtures, examples, docs, datasets or templates are reported as INFO with context: fixture, so sample configs do not fail your CI. Use --include-fixtures to see them at full severity. Only known agent config locations are scanned by default; --all-files scans every JSON and YAML file.

Known limits

• Static and heuristic: it reads files, it does not run servers or check whether a key is live.
• It does not know whether an unpinned package is pinned by your lockfile or a wrapper script.
• AS003, AS005, AS006 and remote-endpoint findings are review prompts, not vulnerabilities.
• Recall has not been measured against a labelled corpus.

Custom rules

Point --rules at a JSON file to add your own org-specific pattern checks without forking the scanner:

agentscan . --rules my-org-rules.json

See agentscan/rules.py (load_pattern_rules) for the schema.

License

MIT. Use it, ship it, fold it into your pipeline.