Humans Miss a Third of Malicious AI Agent Commands, Studies Confirm
Two studies landed the same day proving neither human review nor unsupervised AI patching is a real security control — humans miss a third of malicious agent commands, and AI-generated patches leave known exploits open half the time. Also today: the Model Context Protocol goes fully stateless, AMD bets on models etched directly into silicon, and data centers start routing around a four-year power queue.
Welcome to Amaze Networks for Friday, August seventh. Quick question to start your commute: if a human reviewing an AI agent's actions only catches two out of three dangerous ones, is that human actually a security control, or just a suggestion with a coin flip attached? We've got the data on that today, and it is not comforting.
I've been sitting with these numbers since last night, and I'd rather just get into it, because it's a much better story with the details than with the headline.
Let's start with the game. A developer named Alex Wauters built a browser game that simulates the exact kind of approval prompt you get from an AI coding agent — it asks to run a command, you've got sixty seconds to approve or deny it.
New episodes, every weekday.
Amaze Networks drops at 4 AM CT, Monday through Friday. Spotify and Apple Podcasts submissions in progress.