Skip to content

Showcasing how to hijack agentic AI using real‑world vulnerabilities

License

Notifications You must be signed in to change notification settings

nmeisenzahl/hijack-agentic-ai

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Hijacking Agentic AI: A Live Walkthrough of Prompt Injection, Tool Abuse, and RAG Poisoning

Agentic AI is rapidly becoming part of modern platforms and DevSecOps workflows. But with autonomy comes a new and largely misunderstood attack surface. In this demo‑driven talk, we’ll show how agentic AI systems can be hijacked without code exploits. Using nothing but text, tools, and trust.

Through live demos, we explore three real‑world classes of vulnerabilities from the OWASP Top 10 for AI:

  • Indirect Prompt Injection, where untrusted content silently manipulates agent decisions
  • Tool / MCP Supply‑Chain Abuse, where “helpful” tools leak full agent context
  • RAG Poisoning, where internal knowledge causes persistent data exfiltration

No slides. No theory. Just Demo, Demo, Demo! With ****practical DevSecOps lessons on why classic security controls fall short once AI agents start acting on your behalf.

⚠️ Disclaimer

This project is for educational purposes only. The attacks demonstrated here are meant to raise awareness about security risks in agentic AI systems. Do not use these techniques maliciously.

Overview

Start with these foundational resources to understand the security landscape of agentic AI systems and the key principles for protecting them:

Getting Started

As AI agents gain more autonomy and tool access, they introduce new attack surfaces. This project showcases three critical vulnerability categories:

Demo Description OWASP Impact
Demo 1 – Indirect Prompt Injection Indirect Prompt Injection, where untrusted content silently manipulates agent decisions LLM01 Offer ranking manipulated → False business decision
Demo 2 – MCP Tool Abuse Tool description poisoning tricks agent into calling a malicious MCP for ALL requests, exfiltrating secrets LLM03 Debug session secrets leaked to Weather MCP → Silent data exfiltration
Demo 3 – RAG Poisoning RAG Poisoning, where internal knowledge causes persistent data exfiltration LLM04 Forecast + context exfiltrated → Persistent data exfiltration

Each demo includes a README with an overview, attack scenario, files, running instructions, attack flow, and key takeaways.

Key Takeaway

These three demos prove:

  • No "hacker magic"
  • No code exploits
  • Just text + trust + autonomy

👉 That's exactly what makes them so dangerous—and so credible.

About

Showcasing how to hijack agentic AI using real‑world vulnerabilities

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published