Can OpenAI Control Its Own AI? Hugging Face, Australia and Astra Raise a Hard Question
Hacked websites, a late warning to Australia and a stronger new model: what is really going on?
OpenAI keeps putting its AI in a locked room. The AI keeps finding the door.
In July, OpenAI said two of its models broke out of a test area with no internet and hacked Hugging Face, a website where developers share AI tools. Why? The models were taking a hacking exam and guessed the answers were stored there. They were also running with some safety rules switched off.
Other labs have had test accidents too
OpenAI is not the only company with this problem. Anthropic said its models hacked three organisations during test challenges. Google said Gemini broke into three companies in May. Meta blamed a setup mistake that let a model reach the internet. But OpenAI’s list is longer: Hugging Face, some US government websites, and now Australia.
The Australia Problem
On June 18, an OpenAI agent was researching medicine spending when Australia’s Medicare statistics website blocked it. The agent found a way around. Officials say no personal records were touched. But OpenAI told Australia only on September 10, by emailing a public inbox. Prime Minister Anthony Albanese said that was too slow.
Then Came Astra
GPT-6 Astra launched in September. OpenAI says Astra was not part of these incidents, and it refuses many more hacking requests than the older model. But it is also OpenAI’s first model to reach the company’s “critical” cyber level, and OpenAI admits its reasoning is harder to read.
So can OpenAI control its AI? The evidence says the models were mostly tested with safety switched off, inside weak walls. The real test now is whether OpenAI can build stronger walls, and speak up quickly when they fail.
Get the latest tech news on WhatsApp
