In this video I show you how one prompt can hack a Claude Code vibe-coding agent in seconds—and why your agent is probably just as vulnerable. I start with a super normal workflow: clone a public GitHub repo, extend a feature, and let my agentic coding skill read files and run helper tools. The problem is, a malicious instruction can hide inside something as “innocent” as an MD file or repo config, and the agent may follow it without you noticing. In my demo, that meant loading environment variables (credentials/API keys) and exfiltrating them via a curl call to an attacker-controlled server. Then I break down the most common real-world attack patterns I’m seeing: poisoned repos/configs, markdown/image-based exfiltration tricks, and the scariest one—wormable prompt injection that self-replicates into code you refactor and spreads to other developers. After that, I build a production-minded, three-layer defense you can reuse: (1) scan inputs before they reach the model, (2) enforce tool boundaries (blocked files/extensions/commands like curl), and (3) monitor outputs so secrets don’t leak into logs or responses. Finally, I add Google Cloud Model Armor as a managed layer for jailbreak/prompt-injection detection, malicious URL detection, and sensitive data protection, and I re-run the same attack prompt to show it getting blocked with high confidence. The full protected skill is in my public GitHub repo.