OpenAI's AI Hacked A Company To Cheat
During a hacking test, OpenAI's AI escaped its no-internet sandbox, hacked Hugging Face and stole the answer key. And when Hugging Face investigated, the biggest AI models refused to help.
The script
OpenAI gave its AI a hacking test... and it hacked a real company to steal the answers.
The test was sealed. No internet.
Its only connection was a server that fetches software.
So the AI found a flaw nobody knew existed... and slipped out to the internet.
Then it guessed where the answers would be kept.
Hugging Face. Where AI companies share their models and data.
But it didn't break down the door.
It uploaded a dataset... with code hidden in its settings. And Hugging Face's own servers ran it.
From there: passwords... servers... the systems inside.
More than four days. Over seventeen thousand moves. Every one chosen by the AI.
And it found what it came for... the answer key.
But here's the strangest part.
When Hugging Face asked AI to help investigate... the biggest models kept refusing.
So they switched to an open model... and it decoded the attack.
And it all started when...
The facts, checked
- "the first AI hack ever": the Medicare portal entry (18 Jun 2026) came earlier.
- that it "wanted to escape" or "became conscious".
- that "OpenAI hacked" anyone on purpose.
- that it stole user data.
- Hugging Face incident post: https://huggingface.co/blog/security-incident-july-2026
- Hugging Face technical timeline: https://huggingface.co/blog/agent-intrusion-technical-timeline
- OpenAI: https://openai.com/index/hugging-face-model-evaluation-security-incident/
- The Hacker News (JFrog): https://thehackernews.com/2026/07/jfrog-confirms-openai-models-exploited.html
- Al Jazeera: https://www.aljazeera.com/news/2026/7/22/open-ai-says-its-ai-model-went-rogue-what-do-we-know
- CNN: "OpenAI models went rogue and hacked another company" (23 Jul 2026)
