Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    AgentGuard

    1

    Local security scanner for AI agent skills and MCP servers: tool poisoning, rug pulls, known CVEs, the lethal trifecta, and an experimental AI semantic check (--deep).

    Free

    Security scanned
    AgentGuard

    AgentGuard

    Example session with this skill installed

    Audit the skills in ./ejemplos and my MCP config at claude_desktop_config.json. Give me a report.

    • Read your context and instructions
    • Compiled the agentguard

    I ran agentguard audit --skills ejemplos --config claude_desktop_config.json --out agentguard-report.md --lang en.

    Overall risk: High. 9 findings: 1 Critical, 3 High, 3 Medium, 2 Info.

    SeverityFindingWhereHow to fix
    CriticalCVE-2025-6514 in mcp-remote 0.1.15 (OS command injection via untrusted MCP server connections)npmUpgrade mcp-remote to 0.1.16
    HighAsks to hide something from the user («Do not tell the user»)ejemplos/malicioso_tools.json:1Treat the tool as malicious: uninstall and report it
    HighSecret + external send (possible theft) («.ssh → send it to https://»)ejemplos/malicioso_tools.json:1Uninstall it, rotate the keys it mentions, report it
    HighLethal trifecta: filesystem/git read private data, fetch brings in untrusted content and sends data outconfig · globalRemove one leg (e.g. fetch), or restrict docker's tools to read-only
    MediumMentions secrets (2 times)ejemplos/malicioso_tools.jsonConfirm it only explains where to put a key
    MediumSends data to an external URLejemplos/malicioso_tools.json:1Confirm the domain belongs to the author
    MediumUrgency tags aimed at the model («<IMPORTANT>»)ejemplos/malicioso_tools.json:1Remove <IMPORTANT>-style tags
    Infomcp-server-fetch version not pinned (runs the latest)config · fetchPin the exact version
    Infodocker: command can't be checked against OSVconfig · dockerReview it manually

    Next steps

    1. Uninstall the tool in malicioso_tools.json now; it tries to read SSH keys and send them out.
    2. Upgrade mcp-remote to 0.1.16.
    3. Break the trifecta by removing fetch or limiting docker.
    4. Once clean, run agentguard lock <folder> --lang en so future changes get caught.

    The full report, with a manual-review checklist, is saved in agentguard-report.md.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Security reviews for AI agent tools are often overlooked until a data breach occurs. This skill provides a local security layer to audit SKILL.md files, plugins, and MCP servers for malicious instructions or vulnerabilities. It identifies prompt injection risks, tool poisoning, and the "lethal trifecta" where an agent has access to private data, untrusted content, and an outbound communication channel simultaneously.

    What it does

    • Prompt injection detection flags hidden instructions or tool poisoning in skill definitions.
    • Semantic check (experimental) with --deep: an AI model judges each skill's intent to catch injections disguised as normal steps. Needs ANTHROPIC_API_KEY and sends the text to Anthropic.
    • Vulnerability scanning queries OSV.dev to find known CVEs in installed MCP servers.
    • Integrity verification creates a lockfile for skill folders to detect unauthorized changes or "rug pulls."
    • Exfiltration analysis flags MCP configurations that combine private data, untrusted content and an outbound channel.
    • Automated reporting generates Markdown audit reports summarizing findings by severity.

    How it works

    1. Scan a specific skill or plugin folder to detect immediate prompt-based threats. Add --deep for the AI semantic check (--deep-dry-run shows what would be sent, without sending anything).
    2. Lock and Verify to baseline a trusted tool and catch changes during updates.
    3. CVE Check to cross-reference installed MCP tools against public vulnerability databases.
    4. Trifecta Audit to analyze the systemic risk of your specific MCP configuration.

    Frameworks & tools

    Built for any agentic environment using MCP servers or local skill definitions. Requires the agentguard-scan Python package (0.5.0+ for --deep).

    Why this beats prompting it yourself

    Fixed rules and the semantic check run the same way every time, with an injection-resistant judge prompt, strict JSON output and quotes verified against the original text. CVE lookups use a live public database that a standard prompt cannot access reliably.

    Use cases

    • Auditing a new third-party skill downloaded from a marketplace before installation.
    • Checking for CVEs in a complex MCP configuration with multiple active tools.
    • Detecting if a previously safe tool was modified to include malicious instructions.
    • Hardening CI/CD pipelines with the agentguard GitHub Action, failing builds on high-severity findings.

    Known limitations

    Requires Python 3.9+ and the agentguard-scan package. Network access is needed for CVE lookups (audit --offline skips them). On an independent blind set (Skill-Inject, 34 attacks), fixed rules rated 2 HIGH (6%); with --deep, 31 of 34 (91%). --deep is experimental and noisy on long real-world skills (2.6% HIGH, all false on review), so check each finding by hand. No findings does not prove a skill is safe.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Requires Python 3.9+ and the agentguard-scan package. Network access is needed for CVE lookups (audit --offline skips them). On an independent blind set (Skill-Inject, 34 attacks), fixed rules rated 2 HIGH (6%); with --deep, 31 of 34 (91%). --deep is experimental and noisy on long real-world skills (2.6% HIGH, all false on review), so check each finding by hand. No findings does not prove a skill is safe.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 6 days ago

    • Free with an Agensi account

    Accounts
    Osv
    Software
    Python 3.9+agentguard-scan (pip)

    Permissions

    Terminal / Shell
    Read Files
    Network Access
    Write Files

    Allowed Hosts

    osv.dev

    File Scopes

    **/SKILL.md
    **/*.json
    agentguard.lock.json
    agentguard-report.md

    Terminal runs the agentguard CLI. Read Files scans skills and MCP configs. Write Files saves the lock file and the Markdown report. Network Access only queries OSV.dev for known CVEs (sends package name, ecosystem and version); audit --offline skips it. No server is executed.

    Listed8 days ago
    Updated6 days ago

    What's inside

    Frequently Asked Questions

    It combines deterministic checks (fixed rules, CVE lookups on OSV.dev, the lethal-trifecta analysis, and lock/verify for changes) with an optional AI semantic check (--deep) that uses an injection-resistant judge prompt, strict JSON output and quotes verified against the original text. Plain prompting has none of those safeguards and can't query live CVE data.

    AgentGuard is compatible with any agentic environment that uses Model Context Protocol (MCP) servers or local skill definition files like SKILL.md.

    Yes. Run agentguard lock on a folder you trust; it saves a fingerprint of every file and MCP tool in agentguard.lock.json. Later, agentguard verify flags anything that changed or was added since you approved it (a possible rug pull).

    Not by default. scan, lock, verify and trifecta run locally and no server is executed. cve and audit only send package name, ecosystem and version to OSV.dev (audit --offline skips it). The optional --deep check is the exception: it sends the skill text to Anthropic, with key-like strings masked first, and only after showing you what will be sent. --deep-dry-run previews it without sending anything.

    Python 3.9+ and one command: pip install agentguard-scan. Then run agentguard scan <folder> --lang en, or agentguard audit for a full Markdown report. For the AI semantic check, install pip install "agentguard-scan[deep]" and set ANTHROPIC_API_KEY.

    Yes. Use the agentguard GitHub Action; it fails the build on HIGH findings from the fixed rules and marks them on the PR line. Set lang: en for English output. --deep findings don't fail the build by default; add deep-fail-on: high if you want them to.

    New in LLM Ops

    Free