S3-E14 · Don't Get Pwned (Prompt Injection and the Lethal Trifecta)

You tell an AI assistant connected to your email, calendar, and files to read your inbox and summarize what needs attention, and buried in an unopened message is text written not for you but for the AI. This lecture explains why that is dangerous and why there is no clean fix the way databases have one for SQL injection. You will understand the structural root cause (a model reads its trusted instructions and the untrusted data it fetches as one flat token stream with no boundary), the difference between jailbreaking and indirect prompt injection, the lethal trifecta that turns a helpful agent into a hijacked one, a real responsibly-disclosed incident, and the defenses that genuinely help (cutting capabilities, human-in-the-loop, and the CaMeL architecture). Defender-framed throughout, with patched exploits kept historical. Full course playlist:    • How AI Works · Season 3: Under the Hood   New lecture every week. Subscribe to @HowAIWorksHQ to understand how AI really works, one clear idea at a time.