In medieval courts, a trusted advisor with the wrong loyalties could bring down a kingdom from the inside. LLM-powered apps have the same problem and most teams don't find out until it's too late. You've probably used an LLM API, everyone is building with LLMs right now - either to help them work or integrate into a product. Customer service bots, lifestyle assistants, internal tools with RAG pipelines sitting on top of company data. And most of them have security assumptions baked in that are... optimistic. This talk is about what actually happens when you start poking at these things. I'll walk through what I've found testing real LLM-powered apps: a knowledge base that turned out to give way more access than anyone intended, a lifestyle assistant that was so eager to help it forgot it had limits, and a few bypass techniques that made moderation feel more like a suggestion than a guardrail. One of them took a single prompt and I broke the Cyber Security Agency of Singapore's scammer bot in a way they didn't expect. This isn't a "here's how to be evil" talk — it's a "here's how these things actually fail so we can build them better" talk. We'll look at why LLM security is genuinely different from traditional app security, why prompt injection is a real problem and not just a fun party trick, and what practical steps you can actually take when you're integrating an LLM into a product. If you're building with AI, or you're working with a team that is, this one's for you. Key takeaways: - A tour of the most common ways LLM-integrated apps fail in the real world - Why knowledge bases and overly permissive agents running on your laptop introduce risks most devs don't anticipate - How moderation bypasses work and why they're harder to fix than they look - Practical things you can actually do to make your LLM integration more resilient and a bit of AI ethics sprinkled into the mix