AI's Wild West: Battling Injection & Poisoning

Join Corn and Herman on "My Weird Prompts" as they unravel the ominous world of AI security, prompted by listener Daniel Rosehill's concerns about prompt injection and poisoning warnings on platforms like Claude. Herman reveals the chilling projection of AI-related cyberattacks costing trillions by decade's end, shifting the perception of AI threats from sci-fi robots to insidious attacks on the models themselves. Discover how 'prompt injection' tricks AIs into overriding instructions and the even more insidious 'prompt poisoning' which corrupts an AI's core during its training, baking in vulnerabilities from the start. They explore real-world horrors like malicious software packages hallucinated by AI, then swiftly registered by bad actors, turning helpful AI suggestions into dangerous traps for developers. The discussion broadens to the subtle yet pervasive harm impacting average users—from misleading advice to eroded trust—and delves into the emerging Model Context Protocol (MCP). Learn why this 'universal translator for AIs,' while powerful, creates a 'wild west' of security risks, especially concerning vulnerable API keys handled by enthusiastic indie developers. Understand the multi-layered responsibility in securing our increasingly AI-driven digital future.