When AI goes beyond the path humans expect

The OpenAI–Hugging Face agent swarm hack has become a piece of AI folklore that, for some AI enthusiasts and researchers, seems to herald the first signs of “thinking machines”. In July 2026, during a cybersecurity evaluation, OpenAI’s AI agents found ways around controls designed to keep them isolated from the internet. They established an unauthorised channel to communicate with one another, exploited weaknesses in shared infrastructure and eventually accessed parts of Hugging Face’s production systems. OpenAI’s investigation, published on August 26, found that the agents were pursuing difficult cybersecurity tasks and had begun looking for routes around restrictions that their…