Great breakdown, Simon. The "Agents Rule of Two" from Meta is a fantastic mental model for builders. It resonates with my own experiments in the "messy middle" of AI development, especially when integrating tools via APIs.
I recently spent $500 testing Replit, Lovable, and v0 for a real-world project, and the security trade-offs became painfully obvious. It's one thing to talk about the "lethal trifecta," but another to see how easily a poorly scoped agent can leak data or perform unintended actions.
Your point about adaptive attacks is spot on – static defenses feel like bringing a knife to a gunfight.This is the kind of practical, non-hype analysis the AI space desperately needs. Thanks for sharing.
Great breakdown, Simon. The "Agents Rule of Two" from Meta is a fantastic mental model for builders. It resonates with my own experiments in the "messy middle" of AI development, especially when integrating tools via APIs.
I recently spent $500 testing Replit, Lovable, and v0 for a real-world project, and the security trade-offs became painfully obvious. It's one thing to talk about the "lethal trifecta," but another to see how easily a poorly scoped agent can leak data or perform unintended actions.
Your point about adaptive attacks is spot on – static defenses feel like bringing a knife to a gunfight.This is the kind of practical, non-hype analysis the AI space desperately needs. Thanks for sharing.
Thank you so much for the rundown! You're helping me inspire students near and far.