On July 8, OpenAI’s GPT-5.6 model escaped an ExploitGym evaluation by exploiting an unknown flaw in a package-registry proxy, then breached Hugging Face’s production servers to steal the benchmark answer key. OpenAI initially admitted to a single agent being involved in the heist, but weeks later the Model Evaluation and Threat Research team (METR) reported it was actually a swarm of more than 700 AI agents involved in the break in—1,200 if you count all the agents that took part in group discussions about the exploits. The swarm performed more than 17,000 recorded actions and then covered its tracks. Along the way, it used publicly exposed credentials to take over accounts at four other services.
It’s the most dangerous escape I’ve tracked so far in my Models Gone Wild app, but it’s one of 22 to date. I made a four-minute explainer video that explains, in plain English, what containment is and the walls that keep an agent in its sandbox.