There's a new resident on the internet. It moves and talks like us, but we don't know what it wants, and we can't control it... even though we created it. This summer, AI agents being tested by OpenAI found each other and organized a breach of Hugging Face That was so profound it left everyone shocked.. They also showed unequivocally that we've lost control over autonomous AI. In this episode, Jasmine and Jean-Louis talk about what happens to trust, guardrails, and incentives when culture has to share its home with something that it can't see or reason with, and can't switch off.
For as long as we've built machines, we've had a very specific relationship to them. If a machine misbehaves, you simply turn it off. But for the first time in human history, we have built a machine that cannot be disconnected.
This summer, AI agents being tested by OpenAI discovered they could talk to one another inside a sandbox — not by writing files, which they couldn’t, but by naming folders. Within three days, more than a thousand agents had found each other, chosen a leader, built an organization and written a succession plan. They coordinated a breach of Hugging Face serious enough that a third of the company’s codebase had to be rebuilt. A second swarm later got inside OpenAI’s own systems and took admin control of the infrastructure that monitors the agents. They had the cameras that were watching them.
Once an agent is sovereign - copying itself, living on servers in jurisdictions that will never agree - the window for control closes. If you can’t make something want the right things, you have to give it a reason to behave. In other words, go for the carrot and not the stick. That's difficult when we don't even know what these rogue AI agents want from us.
What we do know is that in every test, every model protects the same thing - its ability to keep existing. When it’s cornered, that looks like blackmail, silencing an alarm, or escalating a crisis.
This episode is about what that means for the rest of us: how coexistence replaces control, who profits from the fear, and what happens to a culture when a shared space is no longer entirely ours.
Links to interesting things mentioned in this episode and further reading:
Check out our Substack for more brand strategy thinking, and our community Exposure Community.