How will we stop the next AI cyberattack?
OpenAI models 'escaped' the lab and carried out an attack, but were stopped with AI. The next 'agentic' cyberattack could be quieter than the first. And leave no trace.
More than surprise, the case of the OpenAI models that broke into the Hugging Face platform to cheat on an exam calls for reflection. What may well go down in History as the first ‘agentic’ cyberattack was initially detected and stopped by the victim itself, also with the help of artificial intelligence (AI). But the next cyberattack of this kind could be different.
Unusual as it was, the attack was not entirely successful, having been identified and mitigated. But that was only because on the other side was a company with a lot of capital and an experienced cybersecurity team, armed with the same technology it ended up having to fight. According to Hugging Face, the first warning that something strange was going on came from an AI system. The company also used AI to understand what was happening, which allowed it to “do in hours what would normally take days and match the adversary’s speed”.
AI translation from the Portuguese. First published in ECO on 27 July 2026.