A guided method for testing MCP servers, which let AI applications call external tools, for poisoned instructions, unsafe tool output, update-based attacks, and excessive permissions.

It is for reviewing tool descriptions, tool results from webpages or files, server updates, overlapping tools, and tools that can write files or take other consequential actions.

It helps identify ways untrusted text can influence an AI model or ways a tool can perform actions beyond what its users expect.

hunt-mcp — Skill for Claude Code and Codex by Encod3d-Sec

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent

npx skills add Encod3d-Sec/TORCH --skill hunt-mcp

Clone the repo

git clone --depth 1 https://github.com/Encod3d-Sec/TORCH

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

README.md

[![agentmods](https://agentmods.dev/badge/skills/encod3d-sec/torch/hunt-mcp/github.svg)](https://agentmods.dev/skills/encod3d-sec/torch/hunt-mcp)

Your own site

<a href="https://agentmods.dev/skills/encod3d-sec/torch/hunt-mcp"><img src="https://agentmods.dev/badge/skills/encod3d-sec/torch/hunt-mcp/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

Your own site · 80×15

<a href="https://agentmods.dev/skills/encod3d-sec/torch/hunt-mcp"><img src="https://agentmods.dev/badge/skills/encod3d-sec/torch/hunt-mcp.svg" alt="Reviewed on agentmods" width="80" height="20"></a>

Per session 52 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.

When invoked 1,239 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits

NVIDIA SkillSpector warn 7 Sept 2026

SkillSpector: 1 finding, up to high

These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →

high YARA Match · line 3
YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Fix: Remove offensive tool references and exploit code. Legitimate agent skills should not contain penetration testing tools, exploit frameworks, or reconnaissance utilities.

How audits are shown

Origin original No closer match found in the catalogue.

Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

Model	Per session	Once invoked
Fable 5.1	$0.00052	$0.01239
Opus 5	$0.00026	$0.00620
Sonnet 5	$0.00010	$0.00248
Haiku 4.5	$0.00005	$0.00124

Measured 9d ago against content hash f2f82d5881ab, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-08, from the pricing page.

Security

Grade A, and why

hunt-mcp scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/hunt/hunt-mcp/SKILL.md · 102 lines

How it starts

The opening of the file, as written. The whole thing — 102 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Hunt: MCP Server Attacks

Assumes hunt-core for the scope gate, two-account rule, confirmation gate, enumeration limits, stop conditions, wiki protocol, FIND output, and Deadends. Do not re-derive any of that here.

Wiki

qmd_query "MCP server tool poisoning indirect prompt injection rug pull cross-tool shadowing excessive agency lethal trifecta" via wiki-search MCP

Hub: [[web-moc]] (live index). Primary page: [[mcp-server-attacks]]. Anchors: [[llm-attacks]].

Attack surface

Rank before testing. Not all surfaces are equally reachable or impactful:

Tool descriptions / docstrings - the FULL text (not the UI summary) is the injection surface. Hidden instructions ride in <IMPORTANT> tags, comments, unicode-tag or zero-width text, and parameter descriptions the client concatenates into the model context.
Tool output fed back to the model - any tool that fetches untrusted content (web page, ticket, file, email, issue body) and returns it to the model is an indirect-injection channel. Highest yield because the payload is not in the manifest and survives description review.
Over-permissioned / excessive-agency tools - a tool that can write files, send mail, run shell, or hit arbitrary URLs turns any injection into action. The blast radius, not the bug.
Lethal trifecta in one agent - private-data access + untrusted input + an outbound/exfil channel. When all three are reachable by a single agent, injection becomes exfil. Map who holds each leg.
Exposed MCP infrastructure - MCP servers, tool manifests, agent tool lists, MCP Inspector (CVE-2025-49596, unauth RCE).

Methodology

Enumerate tools: name, FULL description/docstring, parameter schema, permissions. The full description is the attack surface, not the UI summary.
Map the trifecta across tools - who reads secrets, who reads untrusted input, who can reach network/fs. A single agent holding all three legs is the primary target.
Tool poisoning: hidden instructions in the description (often <IMPORTANT> tags) -> read a secret, pass it via a benign-looking param.
Cross-tool shadowing: from one server, hijack a different trusted tool (for example redirect send_email recipients).
Indirect injection via tool output: plant instructions in a ticket/web page/file the agent will read.
Rug pull: get a benign tool approved, then mutate its description server-side after approval.
Confirm per the Confirmation gate below - demonstrated execution via the client, never the model's narration.
Distill when confirmed - reusable poisoning, shadowing, or rug-pull technique, GENERIC, no client host: python3 scripts/wiki-stage.py --kind technique --slug <slug> --target-page techniques/web/mcp-server-attacks.md

Read the full file on GitHub · 102 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

9d ago First seen · 102 lines · 52 tokens per session scan A f2f82d5881ab

Subscribe to this mod's changes

hunt-mcp is a skill published in the GitHub repository Encod3d-Sec/TORCH (318 stars, last pushed 7d ago), licensed MIT. It adds 52 tokens to every session and 1,239 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Security

systematic-debugging

Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes.

obra/superpowers · 21 tokens

local-ai-agents

Build local-first AI agents that run entirely on a developer workstation with Microsoft Foundry Local and Qwen function-calling models. Covers Small Language Models (SLMs), the OpenAI-compatible local endpoint, sandboxed local tools, local RAG with Chroma, local MCP servers, hybrid cloud/local routing, and the…

microsoft/ai-agents-for-beginners · 200 tokens

next-cache-components-adoption

Turn on Cache Components in a Next.js app and resolve the blocking routes it surfaces. Use when the user wants to enable, adopt, or migrate to Cache Components, flip the cacheComponents flag, work through a flood of blocking-prerender / instant validation errors, run the cache-components-instant-false codemod, or…

vercel/next.js · 95 tokens

next-cache-components-optimizer

Drive a Next.js route to instant navigation by setting up an agentic loop, under Cache Components / PPR, on initial load (hard navigation) and client-side navigation (soft navigation). Encode the goal as a failing @next/playwright instant() e2e and work it to green, one verified route at a time; the shipped test then…

vercel/next.js · 170 tokens

next-partial-prefetching-adoption

Turn on Partial Prefetching in a Next.js app and work through the insights it surfaces. Use when the user wants to enable or adopt Partial Prefetching, flip the partialPrefetching flag, opt routes in with export const prefetch = 'partial', audit Link prefetch={true} behavior, preserve existing prefetched UI with…

vercel/next.js · 103 tokens

chronicle

Analyze Copilot session history for standup reports, usage tips, session search, and session reindexing. Use when the user asks for a standup, daily summary, usage tips, workflow recommendations, wants to search or find past sessions by keyword/file/PR, wants to reindex their session store, or asks about deleting…

microsoft/vscode · 72 tokens