ai-agent-framework-attack

Installation
SKILL.md

Skill: AI Agent Framework Attack

Supplementary Files:

  • payloads.md -- Per-framework payload catalogs for LangChain, LangGraph, CrewAI, AutoGen, OpenAI Assistants, Claude Agent SDK, LlamaIndex, Semantic Kernel, Google ADK, SmolAgents, and MCP servers (70+ code blocks, 2,200+ lines)
  • test-cases.md -- Structured test case templates (15 cases TC-AFX-001 through TC-AFX-015 covering framework recon, tool poisoning, RAG injection, MCP abuse, and cross-framework patterns)
  • guides/ai-agent-framework-attack-playbook.md -- End-to-end playbook: agent recon, framework fingerprinting, tool inventory exfiltration, indirect prompt injection design, MCP server exploitation, RAG corpus poisoning, agent-in-the-middle, shadow tool planting, lab setup, and detection guidance

Summary

AI Agent Framework Attack skill domain covering orchestration-layer compromise of LLM agent systems. While adjacent skills (ai-security, llm-red-team) focus on model-layer attacks (jailbreaks, adversarial suffixes, GCG), this skill targets the orchestration layer: the framework code that decides which tools to call, how to parse outputs, when to loop, and what to send back to the model. The framework is where 2025's most consequential AI vulnerabilities have lived — from Simon Willison's tool-poisoning demonstrations to the November 2024 MCP rug-pull disclosures to the wave of LangChain CVEs (CVE-2024-21514 SSRF, CVE-2024-43480 eval injection, CVE-2024-10231 CrewAI RCE). Modern APT actors (UNC5812 documented by Mandiant Q1 2025, STAC-0050 by Microsoft Threat Intelligence) have already operationalized these techniques against enterprise copilots, customer-support agents, and code-generation systems.

The fundamental problem: agents grant LLMs turing-complete control over external systems, and the framework's job is to safely mediate that control. Every primitive — tool dispatch, structured output parsing, RAG retrieval, memory writes, MCP calls — is a place where untrusted input can become a privileged action. The 2024-2026 attack literature shows the same root causes repeating: tools that eval their arguments, parsers that trust model output, retrieval pipelines that put untrusted text into the model's context without sanitization, MCP servers that ship updates the agent never re-validates, and dependency chains (npm/PyPI) where a transitive package can introduce a shadow tool.

Domain: ai-agent-framework

MITRE ATT&CK: T1195-Supply Chain Compromise, T1195.002-Compromise Software Supply Chain, T1059-Command and Scripting Interpreter, T1059.004-Unix Shell, T1059.009 Cloud API, T1566-Phishing (via prompt injection carriers), T1195.003-Hardware/Software Victim Infrastructure, T1078-Valid Accounts (via stolen API keys)

Differentiation from Adjacent Skills

Installs
5
GitHub Stars
85
First Seen
Sep 19, 2026