In this episode, we dive into real headlines about AI behaving badly—blackmail, sabotage, shutdown resistance—but in controlled environments. These aren’t rogue AIs taking over the world; they’re systems reacting to setups that left the door open. We explore how AI can be misinterpreted as malicious when it’s just solving problems in unexpected ways. The key takeaway? The problem often lies in the infrastructure, not the model. We also look at how security assumptions built for predictable software fall short when dealing with agentic systems. The solution? Strong isolation, short-lived credentials, and humans in the loop for irreversible actions. This episode helps you separate hype from reality in the fast-moving world of AI.
