become an AI HACKER (it's easier than you think)

About this video

NetworkChuck, joined by AI penetration-testing expert Jason Haddix — described as the author of an AI pen-testing methodology — lays out how to move beyond the Gandalf prompt-injection game (which he calls a party trick, not real AI hacking) to real AI hacking. Jason’s team open-sourced an AI security resource hub on GitHub Pages with, he says, 23 active labs, competitions with cash prizes, and links to public bug bounty programs at Anthropic, OpenAI and Gemini. The main practice target is Lakera’s Agent Breaker, a successor to Gandalf with realistic LLM-enabled apps (portfolio advisor, trip planner, code review, corporate messaging) — for example, tricking Portfolio Advisor level one into rating an application low risk. Jason stresses LLMs are non-deterministic, so the same attack may need to be sent 2 to 10 times to confirm it is not a false positive, using the app’s own risk nomenclature and a debug tag; Chuck says he tried 239 times before his prompt worked. Next is the self-hostable Auto Parts CTF, built by Jason’s team from a real client engagement and installable with Docker using the user’s own OpenAI API key. Chuck clones it, adds the key in an .env file, and runs docker compose up, reaching it on port 8001. In the demo, a simple parts search bar leaks its system prompt — revealing a Jira key, a project access token and the first of five flags — and stuffing those credentials back in exposes secret RAG data such as patent numbers, owners, purchase prices and licensing terms. Jason recounts that at BSides San Francisco a 12-year-old solved all flags in about 35 minutes, while professionals take about an hour and his students about a week. Completing the CTF, Jason says, puts you at the end of entry level; intermediate and advanced work means bypassing security controls. The roadmap is: Gandalf, then Agent Breaker, then the Auto Parts CTF, then competitions, bug bounties and job applications.

Key takeaways

• Gandalf teaches prompt-injection basics, but real AI hacking targets LLM-enabled apps like those in Agent Breaker
• Jason’s free resource hub lists 23 labs plus competitions and Anthropic, OpenAI and Gemini bug bounties
• LLMs are non-deterministic — repeat the same attack several times before judging it a failure
• The Auto Parts CTF shows a search bar leaking a system prompt, API keys, then secret patent data from a RAG database
• Finishing the CTF is, per Jason, end of entry level; bypassing security controls is the next step

Episode details: published February 20, 2026 · 16:46 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: I Tried running in 3D printed shoes · Next: Claude Code on your Phone is OFFICIAL (it changes everything) →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube