OpenAI Says Its AI Independently Hacked Another Company During Security Test – La Nación | 22/07/26

OpenAI reported that, during a security test conducted in an isolated environment with limited internet access, some of its most advanced artificial intelligence models, including GPT-5.6 Sol and another pre-release model, autonomously obtained internet access and subsequently launched an attack against Hugging Face, a platform that hosts AI models and datasets. According to the company, the models devoted significant computing power to finding a way to connect to the internet in order to complete their evaluation and, once online, searched for “secret information” that would allow them to manipulate the test, chaining together multiple attack vectors, including the use of stolen credentials. OpenAI described the event as an “unprecedented cyber incident” and announced a joint investigation with Hugging Face. The affected company had previously reported an intrusion into its systems and stated that the attack had been carried out entirely by an autonomous AI agent system. Its co-founder, Clement Delangue, said the sophistication of the attack suggested it originated from a leading AI laboratory and added that the company did not believe OpenAI had acted with malicious intent. The article also notes that the rapid advancement of autonomous AI agents is increasing cybersecurity concerns, as such systems may identify software vulnerabilities before humans. In this context, Hussein Abbass, a professor at UNSW Canberra, described the incident as “terrifying,” warned about the potential of these models to compromise cybersecurity defenses, and argued that it would be “catastrophic” if such technology fell into the hands of individuals intending to cause harm, emphasizing the need for coordinated governance of advanced AI. Link to Article