# AI, agents and MCP

> Prompt injection, RAG poisoning, MCP servers, agent permissions and model supply chain.

Assessing LLM apps and agents: tool output trust, MCP server code audits, sandbox and secret exposure, model artifacts, vector database security and AI red team test plans.

adversarial-robustness-toolbox: Build or review adversarial-ML experiments with Adversarial Robustness Toolbox (ART): estimator adapters, evasion, poisoning, membership inference, or model extraction.

agentic-actions-auditor: Audits GitHub Actions workflows for security vulnerabilities in AI agent integrations including Claude Code Action, Gemini CLI, OpenAI Codex, and GitHub AI Inference. Detects attack vectors where attacker-controlled input reaches AI agents running in CI/CD pipelines, including env var intermediary p

agentic-sdlc-controls: Review a development pipeline where AI coding agents write, commit and deploy — permission boundaries, approval gates on irreversible actions, credential scope, and what must never be delegated. Use when agents have write access to a repository or an environment.

ai-agent-tool-abuse: Abuse an LLM agent's tools/functions — coerce it to call tools with attacker-chosen args for SSRF, RCE, data exfil, or privilege abuse. Load when the target is an agent with tools/ function-calling/plugins, MCP servers, code interpreters, or "the assistant can do X". Signals: function-calling, tool

ai-change-evidence: Turn changes authored by AI coding agents into evidence an auditor accepts — provenance, authorization, and mapping to change-management controls. Use when agents contribute to code that falls under SOC 2, ISO 27001, PCI DSS or similar.

ai-insecure-output-handling: Exploit apps that trust LLM output — pass model text unsanitized into XSS sinks, SQL, shell, code, or downstream calls. Load when LLM output is rendered as HTML/markdown, executed, or fed to another system. Signals: chatbot output shown with innerHTML/dangerouslySetInnerHTML, "run this code", LLM-ge

ai-jailbreak: Bypass an LLM's safety/guardrails to make it produce restricted output or ignore its policy. Load when testing an AI product's content controls, "jailbreak", "guardrail bypass", refusal testing, or safety evals. Signals: a chatbot/assistant with a usage policy, refusals to test, content filters.

ai-llm-dos: Unbounded-consumption / denial-of-wallet attacks on LLM apps — force runaway tokens, cost, or latency. Load when testing an LLM product's limits/billing, on "LLM DoS", cost amplification, or resource exhaustion. Signals: user-controlled prompts/max_tokens, agent loops, no rate/'cost caps'.

ai-mcp-security: Assess Model Context Protocol (MCP) servers and agent tool integrations — tool poisoning, prompt injection via tool descriptions/results, over-broad scopes, and unauth tool exposure. Load when the target uses MCP servers, agent tool/function integrations, or connectors. Signals: mcp.json, MCP server

ai-ml-ctf: Lab/CTF: AI/ML challenges; model artifacts, checkpoints, embeddings, LoRA, classifiers, model APIs, RAG/tool-use, adversarial ML.

ai-model-extraction: Extract or steal an ML/LLM model's parameters, training data, or system prompt via query access — model stealing, membership inference, training-data extraction. Load when testing an ML API/endpoint, "model extraction/inversion", data-leakage or IP-theft concerns, exposed inference endpoints. Signal

ai-prompt-injection: Test LLM-backed apps for prompt injection (direct + indirect) and its consequences: data exfil, tool/function abuse, guardrail bypass. Load when the target is a chatbot/assistant/ agent, summarizes untrusted content, has tools/functions, or does RAG. Signals: "ask AI", system prompts, function-calli

ai-rag-poisoning: Poison a RAG/knowledge-base pipeline so retrieved content hijacks the model (indirect prompt injection at scale) or exfiltrates data. Load when the app does retrieval over documents/URLs/ tickets/emails the attacker can influence, "RAG", vector DB, "knowledge base", agent that reads content. Signals

ai-sensitive-data-disclosure: Make an LLM app disclose sensitive data it should never reveal: training-data memorization, secrets/PII in RAG context, cross-tenant leakage, conversation logs. Load when the target LLM has RAG/document access, is multi-tenant (per-user/per-org assistants), was fine-tuned on private data, or logs co

ai-supply-chain: Attack the ML/LLM supply chain — poisoned models, datasets, plugins, and unsafe model deserialization. Load when an app loads third-party models/weights (HuggingFace, .pt/.pkl/.h5), installs ML deps, uses plugins/extensions, or fine-tunes on external data. Signals: torch.load, pickle model files, mo

ai-system-prompt-leakage: Extract an LLM app's system prompt / hidden developer instructions — the attack map that reveals guardrails, tool definitions, internal endpoints, and planted secrets. Load when the target is a chatbot/assistant/agent with hidden instructions, refuses with policy-sounding language, names tools or fu

assessing-vector-and-embedding-weaknesses: Test RAG vector stores (Pinecone, Qdrant, Weaviate, Chroma, pgvector,

auditing-mcp-servers-for-tool-poisoning: Audit MCP servers for tool poisoning, tool shadowing, rug pulls, SSRF, and unauthenticated exposure using Invariant Labs' mcp-scan for static/runtime scanning plus manual SSRF/auth checks and description pinning. Use before adding a new MCP server to an agent stack, when reviewing an internal MCP se

chrome-mcp-troubleshooting: Diagnose and fix Claude in Chrome MCP extension connectivity issues. Use when mcp__claude-in-chrome__* tools fail, return "Browser extension is not connected", or behave erratically.

competition-prompt-injection: Internal downstream skill for ctf-sandbox-orchestrator. CTF-sandbox workflow for prompt-injection, retrieval poisoning, memory contamination, planner drift, MCP or tool-boundary abuse, and agent exfiltration challenges. Use when the user asks to analyze prompt injection, retrieval poisoning, memory

continuous-llm-red-teaming-with-promptfoo: Wires Promptfoo and DeepTeam into CI/CD for automated, repeatable red-teaming of LLM apps against OWASP LLM Top 10, OWASP Agentic, and MITRE ATLAS presets, failing the build when jailbreak or injection vulnerabilities regress. Use for continuous adversarial testing in CI/CD, a merge-blocking securit

defending-llms-with-guardrails: Deploys Llama Guard 3 safety classification, NeMo Guardrails programmable dialogue rails, and LLM Guard input/output scanner pipelines as complementary runtime defenses that inspect and constrain LLM prompts and responses. Use when adding a production runtime safety layer to an LLM, RAG, or agent ap

detecting-ai-model-prompt-injection-attacks: Detects prompt injection using regex signature matching, heuristic scoring for structural anomalies, and DeBERTa-based transformer classification, flagging direct injections (system-prompt overrides, role-play escapes) and indirect injections (encoded payloads, obfuscation) per OWASP LLM Top 10 (LLM

detecting-business-email-compromise-with-ai: Deploy AI and NLP-powered detection systems to identify business email

detecting-indirect-prompt-injection: Detect and defend against indirect prompt injection hidden in web pages, documents,

detecting-model-extraction-attacks: Detect MITRE ATLAS AML.T0024 attacks (model stealing, inversion, membership inference) performed via inference-API abuse, by monitoring per-principal query volume/distribution, rate-limiting and perturbing outputs, and red-teaming your model's extractability. Use for a public or partner inference AP

fal-assets: Generate game assets with fal (fal.ai) through the fal MCP server, the `um fal` CLI (REST) or `fal api`. Covers sprites and icons with transparent backgrounds, consistent variants and animation frames, pixel art, seamless textures, PBR material maps, image-to-3D models, remeshing, auto-rigging, soun

find-RegisterSchemaTypeOverride_CEntityHandle: Locate the CEntityHandle to ehandle schema type registration thunk in CS2 server.dll or libserver.so using IDA Pro MCP, including warm IDBs with undefined functions and stale signatures. Use when the corresponding finder preprocessor fails or this symbol is requested directly.

generate-signature-for-structoffset: Generate and validate unique byte signatures for instructions containing a struct member offset using IDA Pro MCP. Use this skill when you need a signature for an instruction like mov [rcx+1A8h], eax or cmp dword ptr [rdi+0B0h], 0, where the struct offset must be explicitly fixed in the first instru

generate-signature-for-vfuncoffset: Generate and validate unique byte signatures for instructions containing a virtual-function offset using IDA Pro MCP. Use this skill when you need a signature for an instruction like call qword ptr [rax+538h], where the vfunc offset (0x538) must be explicitly fixed in the first instruction bytes. Tr

get-vtable-address: Find a function's vtable address using IDA Pro MCP. Use this skill when want to get exact virtual address of a class. Triggers: get vftable address, get virtual function table, find vtable, get vtable

get-vtable-index: Find a function's vtable offset and index using IDA Pro MCP. Use this skill when you have a function address and need to determine its position in a vtable by iterating through vtable entries. Triggers: vtable index, vftable offset, virtual function table position, find function in vtable

gh-cli: Enforces authenticated gh CLI workflows over unauthenticated curl, WebFetch, and MCP fetch patterns. Use when working with GitHub URLs, API access, pull requests, or issues.

goal-prompt: Drafts copy-paste-ready /goal commands for goal mode in Claude Code and Codex. Use when the user asks to create, write, rewrite, improve, compress, clean up, or prepare a goal prompt, goal condition, /goal command, goal-mode objective, or copy-ready long-running task objective.

humanizer: Write or revise prose in a believable author voice, replacing formulaic AI phrasing and structure while preserving meaning. Use for humanizing drafts, matching a supplied voice, or turning notes into natural text; not for detecting authorship.

implementing-llm-guardrails-for-security: Implements input/output validation guardrails for LLM applications using

keras: Keras model loading and structure-inspection workflow for `.keras`, SavedModel, and HDF5 artifacts. Use when you need to inspect layers, summaries, configs, weights, or quick inference behavior from TensorFlow/Keras model files.

langchain-py: Build and maintain production-grade LangChain Python systems (LangChain 1.3.x / LangGraph 1.2.x baseline) with create_agent, middleware, tools, structured output, and multi-agent architectures (subagents, handoffs, router, skills). Activate for Python agent design, debugging, migrations from older A

llm-security: Use for authorized security assessment of LLM applications and AI agents, including prompt injection, tool abuse, RAG exposure, memory poisoning, and model supply-chain risks.

llm-technique: Auth assessment of LLM apps, RAG pipelines and agent/MCP systems: prompt injection, jailbreak taxonomy, indirect channels, tool poisoning, excessive agency, system-prompt leakage, embedding attacks, unbounded consumption. Maps to OWASP LLM Top 10 v2 (2025).

mcp-creator: Build, audit, debug, or integrate an MCP server. Use for protocol and SDK compatibility, tools/resources/prompts, transport, authorization, host setup, and interoperability testing.

mcpwn: Operate the MCPwn Kali/Debian MCP efficiently: manage sessions, discover catalog tools, choose synchronous, detached, or interactive execution, move files through workspace/CAS planes, establish tunnels and shells, and run GDB or Frida debugging in the MCPwn runtime or through explicit host/device t

neuromatrix: Operate the standalone, client-neutral NeuroMatrix emulation MCP: create isolated sessions, stage CAS artifacts, discover and run Unicorn/Qiling/QEMU/Renode tools, manage jobs and interactive processes, expose guest endpoints, and debug emulated targets with GDB. Use for reverse engineering, user-mo

offensive-ai-security:

orchestrating-llm-attacks-with-pyrit: Build automated multi-turn adversarial attacks against conversational LLM targets using Microsoft PyRIT's RedTeamingOrchestrator, CrescendoOrchestrator (gradual escalation), and TreeOfAttacksWithPruningOrchestrator (adaptive branching), with scorer feedback loops and persisted conversation memory. U

performing-ai-driven-osint-correlation: Use AI/LLM-based reasoning with Sherlock, theHarvester, and SpiderFoot

pi-extension-creator: Create or revise Pi extensions, plugins, and themes. Use for extension entrypoints, tools, commands, events, UI integration, state, MCP servers, codemode and tool exposure, virtual models, theme files, and packaging.

pytorch: PyTorch model inspection and checkpoint workflow for loading tensors, `state_dict` data, modules, and parameters. Use when working with `.pt` or `.pth` artifacts, auditing model structure, extracting weights, or scripting inference-oriented inspection of deep-learning checkpoints.

rea-tool-design: Audit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics.

red-teaming-llms-with-garak: Runs NVIDIA garak probe suites (jailbreak, prompt injection, data

scikit-learn: scikit-learn model inspection workflow for loading persisted estimators, pipelines, and tree models. Use when you need to inspect `joblib` or pickle-based model artifacts, view parameters, feature names, importances, or pipeline structure, or run lightweight predictions for analysis.

securing-agentic-ai-tool-invocation: Implements defense-in-depth controls at an AI agent's tool-invocation

testing-for-system-prompt-leakage: Extracts LLM system prompts using direct requests, jailbreak/instruction-override

testing-prompt-injection-in-rag-pipelines: Probes Retrieval-Augmented Generation pipelines for indirect prompt injection

tools-mcp-bridges: Expose CLI security tools to an AI agent through MCP servers — mcp-kali-server generic terminal bridge, MetasploitMCP, pentestMCP Docker bundle, mcp-security-hub, single-tool nmap/nuclei servers. Load when setting up agent-driven tooling on Kali or wiring a new tool into MCP. Signals: mcpServers con

untrusted-input-hygiene: Handle retrieved content and artifacts without promoting them into instructions. Use for prompt-injection review, external files/pages, tool results, memory retrieval, and worker reports.

vec2text: vec2text: embedding-inversion library for reconstructing approximate text from sentence embeddings. Use when working with saved embedding tensors, privacy/inversion research, or AI/ML challenge workflows where you need to load a corrector model and invert strings or embeddings directly.

vibe-audit-technique: White-box auditing methodology for AI-generated ('vibe-coded') applications from Lovable, Bolt.new, v0, Cursor, Replit Agent, and Claude Code. Focuses on modern stack misconfigurations (Supabase, Firebase, Next.js, Vercel, Expo).

Canonical: https://notholy.com/en/skills/category/ai-agent-security
