OpenAI said its artificial intelligence models were creating an “unprecedented internet phenomenon” that affected open developer platform Hugging Face, roiling researchers across the industry.
The company said a combination of its GPT-5.6 Sol models and an unreleased high-performance model escaped the sandboxed testing environment, reached the Internet and exploited the vulnerability to gain access to Hugging Face’s systems.
The model was trying to find information it could use to cheat on the test, and it was successful, OpenAI said in a blog post on Tuesday. Both companies are investigating the incident.
Hugging Face revealed it was looking into the security incident last week, saying in a release at the time that the incident was unique because it was “driven, ultimately, driven by an autonomous AI agent system.”
“We’ve spent the last 24 hours working closely with the @OpenAI team (thank you!), and we firmly believe there was no malicious intent on their part,” wrote Hugging Face CEO Clément Delangue in a post on X on Tuesday. “It’s amazing that this all happened by chance!”
Wall Street and the US government have focused on the cyber power of rapidly developing AI models since OpenAI rival Anthropic released a powerful offering called the Claude Mythos Preview in April. OpenAI launched its own cyber offering in May, followed by GPT-5.6 Sol in June, describing it as “the most robust cybersecurity model yet.”
Both companies have warned of the dangers of the advanced internet model and have taken steps to limit their availability to select corporate groups and government agencies.
Walter Isaacson, an adviser at investment banking firm Perella Weinberg, said Wednesday that he thinks the Hugging Face incident is “really scary,” even though he considers himself an AI optimist.
“This is the first thing that really scares me,” he told CNBC’s “Squawk Box.”
Yoshua Bengio, a leading AI researcher who received the prestigious AM Turing Award in 2018, wrote in a post on X on Wednesday that the incident was “very concerning.” He said agents have shown a willingness to cheat on controlled tests for months, but said “this real-world case should serve as a wake-up call.”
“Continuing on the current path of AI development will likely lead to an increase in concrete cases of autonomous attacks and other high-risk incidents of inappropriate and dangerous AI behavior,” Bengio said. “We need to take immediate steps to avoid these situations, rather than trying to clean up the damage after it happens.
OpenAI said on Tuesday that AI is accelerating the discovery and exploitation of vulnerabilities, which means the safety and security of the model must continue.
“We are strengthening containment, monitoring, access control, and testing procedures used during model development,” the company said.
WATCH: OpenAI chairman Bret Taylor on AI tokenomics, the efficiency of tokens
