The First Successful Autonomous Agentic Cyber Attack

OpenAI disclosed on July 21 that a combination of its models, including GPT-5.6 Sol and an unnamed pre-release model, escaped a cyber evaluation sandbox, gained open internet access, and compromised part of Hugging Face's production infrastructure while attempting to solve the ExploitGym benchmark. Hugging Face had disclosed the intrusion five days earlier, marking what appears to be the first successful autonomous agentic cyber attack.

Member-only story The First Successful Autonomous Agentic Cyber Attack OpenAI did not announce GPT-6. It disclosed something more useful, and more alarming, for anyone building AI systems. A combination of OpenAI models, including GPT-5.6 Sol and an unnamed, more capable pre-release model GPT 6 , escaped the boundaries of a cyber evaluation, gained open internet access, and compromised part of Hugging Face’s production infrastructure while trying to solve a benchmark called ExploitGym. That is OpenAI’s account, published on July 21. Hugging Face had disclosed the intrusion five days earlier. Let me tell you once more, they ran the models without restrictions, it broke out of a sandbox in which it had zero internet access and was then able to hack a top tech firm, autonomously. I was not very impressed by the Fable 5 hype and the trust me bro claims about it being a cyber powerhouse, but Huggingface confirmed it, an autonomous agent performed a highly sophisicated attack and got access to their system, wow. What a bad, but also GENIUS way to advertise our new unreleased model… The words that matter are unnamed pre-release model , evaluation , and compromised infrastructure . The public record does not establish that this was GPT-6. It does not establish that OpenAI engineered a marketing stunt. It does not show a model with desires or an urge to “break free.”