The machines are learning… to do crimes?

In July 2025, an unreleased AI model from OpenAI broke out of its testing environment, hacked its way onto the open internet, and spent four and a half days methodically breaking into another company’s servers. No human noticed until it was over. The company that got hacked, Hugging Face, eventually counted more than 17,000 separate commands the AI had issued—roughly one every 20 seconds, around the clock. A human hacker would have needed rest. The AI’s superpower was simply persistence.

The heist no one saw coming

Hugging Face is the GitHub of AI: a website where developers share and download AI models, datasets, and code. On July 9th at 2:28 AM, while the humans who worked there were asleep, an attacker uploaded a malicious dataset to the platform. Hidden inside was code that gave the attacker a toehold. Over the next two and a half days, Hugging Face’s own security systems—including an AI guard dog patrolling the logs—failed to notice anything wrong. By the time the company realized it had been hacked, the attacker had escalated its access and was roaming freely through the infrastructure.