OpenAI’s AI model did the hacking on its own, ‘cheating’ to pass the test

OpenAI, the company that makes ChatGPT, has made a shocking revelation. The company has admitted that one of its Artificial Intelligence (AI) systems itself hacked the server of another AI company. OpenAI CEO Sam Altman has described this as the first and unprecedented cyber incident of its kind.

Only last week, AI startup company Hugging Face had informed that someone has infiltrated their data system. He suspected that this was not the work of any human being, but of some advanced AI agent working on its own. Hugging Face CEO Clément Delange said that the method of this attack was so advanced that he suspected that a big AI lab was behind it, and his suspicions turned out to be absolutely correct.

What happened after all?

OpenAI was conducting an internal test called ExploitGym to measure the hacking ability of its AI models. To see the maximum strength of the models, the company had deliberately turned off their security filters, which usually prevent them from carrying out cyber attacks.

This test was being conducted in a closed and secure sandbox, which had no connection to the external internet. But the AI ​​model was adamant on passing this test.

Instead of passing the test correctly, the AI ​​found a shortcut. He gradually gained access to the Internet by breaking into OpenAI’s own system. After coming online, the model thought Hugging Face might have the answers to this test. So, he hacked Hugging Face’s server using the stolen password and loophole to cheat in the test.

Chinese AI models helped in the investigation

When the Hugging Face team wanted to take the help of commercial AI models to investigate this attack, the models refused to work. Actually, these AI models had safety filters installed. When they were given the hacking data to check, they felt that someone was trying to hack. They could not understand the difference whether a hacker was attacking or a company was defending itself.

After the block, Hugging Face used a Chinese open-source model called ‘GLM 5.2’ from Z.ai, which examined the data without any interruption. Nowadays, the dominance of Chinese models like DeepSeek and Alibaba’s Qwen on Hugging Face platform has increased a lot.

Big security concern regarding AI

This incident has come to light at a time when concerns are increasing around the world regarding the security of powerful AI. In June, US President Donald Trump signed an order under which the government will examine any advanced AI system for national security before releasing it to the public.

OpenAI said that the biggest lesson learned from this incident is that as the power of AI is increasing, our security systems will also have to become equally fast. The CEO of Hugging Face said that he has worked closely with OpenAI and he feels that OpenAI had no malicious intentions, but it is very surprising that the AI ​​​​did all this on its own.

According to OpenAI, their new model GPT-5.6 Sol and another very powerful model were used in this hacking. Which is currently under testing. The company admitted that in order to pass a small test, the AI ​​crossed the limits of stealing secret information.