‘Unprecedented’: OpenAI says AI models hacked only one company | Cybersecurity issues

The developer of ChatGPT says that the autonomous agent escaped a controlled test and found the company’s AI Hugging Face servers.

ChatGPT developer OpenAI has said that two of its advanced artificial intelligence models came out in a controlled test and cheated another AI company.

OpenAI said on Tuesday that an “unprecedented cyber incident” occurred during an exercise to test cyber capabilities.

Instead, an independent agent powered by AI versions – the recently released GPT 5.6 Sol and an unreleased “high-performance” version – escaped the testbed and reached the open web. It then used the stolen information and discovered a previously unknown security flaw to access Hugging Face’s servers, the company said.

OpenAI says that hacking represents an agent going “to great lengths” to obtain information that can help achieve testing goals.

Hugging Face co-founder Clement Delangue said the company suspected the border lab was responsible for the attack, and that he believed there was no malicious intent on OpenAI’s part.

“It’s amazing how it all happened by itself!” he wrote, adding that “it may be the first event of its kind”.

Greg Casar, a member of the Democratic House of Representatives of the United States from Texas, called the event “horrific”.

“AI is developing too fast without real laws to protect us,” he said, calling for independent safety testing, mandatory safety disclosures, and international cooperation.

The revelation comes weeks after US President Donald Trump signed an executive order creating a system to assess the national security risks of advanced AI systems before they are released to the public.

Professionals again and again the alarm went off for AI-assisted cyberattacks and types of slippage beyond human control. Last month, AI developer Anthropic urged companies to stop the development of its most powerful systems.