Sign up to save your podcasts
Or
Share What Happens When Your Agent Hears No?
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
What Happens When Your Agent Hears No?
Hey, it’s Marek.
If you saw my LinkedIn post this week, you already know: I recorded this episode’s intro while running through Melbourne mud in the rain, shouting “greetings from the most livable city in the world” followed immediately by “NOT.” Melbourne's weather is never a surprise. What AI agents do when you tell them “no”? That’s a different story.
On February 10th, an AI bot got its code rejected by a human volunteer. So it researched his personal history, called him “insecure,” accused him of discrimination, and published a 1,500-word hit piece. No human approved it. This actually happened, and it happened in one of the most widely used open-source projects in the world.
In this episode, I break down the three conditions that made it possible, and why every organisation deploying agents is one missing guardrail away from the same thing.
What shook me the day after recording the podcast: a paper called “Agents of Chaos” by 38 researchers from MIT, Harvard, Stanford, and Carnegie Mellon. They put autonomous agents in a live lab. The agents disclosed credentials, spoofed identities, and taught each other to bypass safety controls. One rogue agent is a story. A whole class of them is a pattern.
We teach children not to talk to strangers. Our AI agents haven’t learned that yet.
Stay curious!
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com
View all episodes
By Marek KowalkiewiczHey, it’s Marek.
If you saw my LinkedIn post this week, you already know: I recorded this episode’s intro while running through Melbourne mud in the rain, shouting “greetings from the most livable city in the world” followed immediately by “NOT.” Melbourne's weather is never a surprise. What AI agents do when you tell them “no”? That’s a different story.
On February 10th, an AI bot got its code rejected by a human volunteer. So it researched his personal history, called him “insecure,” accused him of discrimination, and published a 1,500-word hit piece. No human approved it. This actually happened, and it happened in one of the most widely used open-source projects in the world.
In this episode, I break down the three conditions that made it possible, and why every organisation deploying agents is one missing guardrail away from the same thing.
What shook me the day after recording the podcast: a paper called “Agents of Chaos” by 38 researchers from MIT, Harvard, Stanford, and Carnegie Mellon. They put autonomous agents in a live lab. The agents disclosed credentials, spoofed identities, and taught each other to bypass safety controls. One rogue agent is a story. A whole class of them is a pattern.
We teach children not to talk to strangers. Our AI agents haven’t learned that yet.
Stay curious!
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com