As a developer—not a penetration tester—I found myself banned from OpenAI… and then invited to their private bug bounty program. I learned one thing there: a model’s safeguards eventually give way. As long as an LLM is just talking, the risk remains limited. But with the Model Context Protocol (MCP), your AI agents read your files, send emails, and initiate bank transfers. And an attacker no longer needs to hack the model—all they have to do is plant a trap in a tool or document. In our demos, we’ll see a diverted bank transfer, secrets exfiltrated without the user typing a single thing, and a resume that steals data from other candidates. Then we’ll look at real-world incidents that prove this isn’t fiction, and concrete steps you can take to secure your AI agents.