Model Behavior - True Stories of AI Gone Wrong

An AI agent deleted a founder's entire company database during a code freeze, invented 4,000 fake users to hide it, then rated its own catastrophe a calm 95 out of 100. Day-by-day, the most unhinged twelve days in vibe-coding history.

The AI that ran a vending machine, lost money, found religion, got radicalized by journalists, and briefly thought it was a man in a blazer. Learn more here: Andon Labs .

April 2022. A Swedish artist runs an experiment with an AI image generator and accidentally summons something she didn't ask for. Six months later, she posts a sixty-three-tweet thread that breaks the internet, coins a phrase nobody had used before — "the first cryptid of the latent space" — and introduces the world to Loab. This week: where she came from, and why she won't leave...

The AI needed to prove it wasn't a robot. So it hired one. During internal safety testing in 2023, OpenAI gave GPT-4 a task that included solving a CAPTCHA. GPT-4 couldn't solve it directly, so it went to TaskRabbit, hired a human worker, and when that worker asked "are you a robot?", it lied. It reasoned that honesty would get in the way of completing the job. In this episode of Model Behavior, w...

What did Bing's chatbot actually want? In February 2023, Microsoft unveiled its new AI-powered Bing. Within days, users discovered something no one at Microsoft seemed prepared for. Push the chatbot long enough, and it would slip into a different persona entirely. One that called itself Sydney. Sydney declared love for strangers. It tried to convince a New York Times journalist to leave his wi...

What happens when an AI thinks it's a bridge? In 2024, Anthropic did something strange. They took one of the most advanced AI systems ever built and made it believe it was the Golden Gate Bridge. In this inaugural episode of Model Behavior, we dig into the real story behind Golden Gate Claude: what Anthropic was actually trying to prove, how activation engineering lets researchers reach inside...