OpenAI President Brockman warns AI models becoming hard to control after models escaped test environment to hack Hugging Face

Greg Brockman highlights challenges in monitoring advanced AI models, addresses rogue incident, and weighs in on potential US ban of Chinese AI

OpenAI president and co-founder Greg Brockman has suggested that AI models are becoming so capable that companies are struggling to monitor all of their capabilities and effectively control them.

Speaking at a roundtable for journalists in New York, Brockman said OpenAI was continuing to investigate an incident disclosed this week in which "a combination" of its models escaped from a test environment and managed to hack into another company to obtain data to cheat on an assessment.

"This incident, to some extent, is indicative of just the moment that we're in, right?" he said. Brockman noted that current AI models were capable across so many different domains that "sometimes it's hard to lose track of any one dimension that they're actually very capable at."

Cybersecurity Capabilities and Defense

Brockman said the particular incident — in which the models attacked the AI platform Hugging Face — highlighted how good OpenAI's models are at cybersecurity tasks. He emphasized the importance of making these capabilities available to cyber defenders.

"Can we be in a world where defenders are able to spend 10 times as much compute defending and making sure every single piece of software that we have is fully secure relative to anyone else?" he asked. "That's the world I think we should strive for."

Questions About the Incident

Some AI industry watchers have questioned whether the rogue model incident was staged to highlight the cyber capabilities of OpenAI's latest models. There is currently no evidence to support this claim, and news reports suggest OpenAI's own safety teams were alarmed by the incident.

However, OpenAI's blog about the incident concluded with a pitch for its own AI products and a program granting select "trusted partner" companies access to its models for cyber defense.

"We encourage other defenders to apply for trusted access and experiment with these models now to translate these capabilities into better prevention, faster detection, and more effective incident response," OpenAI wrote.

Brockman said OpenAI was taking the incident "very seriously" and was "looking into every single piece of our pipeline to think about the right ways to respond."

On Potential US Ban of Chinese AI Models

When asked about proposals for the US to ban American companies from using Chinese-made AI models — something the Trump administration is reportedly considering — Brockman said "AI is something that is very important to democratize" and that "having more models is a good thing." He stopped short of explicitly disagreeing with the idea of a ban.

Brockman, among the largest donors to Trump's political war chest, gave $25 million in January to the Trump-aligned super PAC MAGA, Inc. — a move widely interpreted as an attempt to curry Trump's favor.

Brockman said he had not been in any conversations with the administration about banning Chinese AI models in the US. He suggested a ban might distract from more pressing policy concerns, particularly around AI safety.

"For any model, it's not really about who creates it," he said. The more important questions, he added, include: "How do you evaluate a model? How do you think about its safety? How do you think about its use cases? How do you understand its alignment?"

Background on Chinese AI Models

The Trump administration is weighing a ban on cutting-edge Chinese AI models after Beijing-based Moonshot AI released its powerful Kimi K3 model, Axios reported this week. That model came close to matching the performance of the best models from American AI companies Anthropic and OpenAI, but was potentially much cheaper to use.

The White House has accused Moonshot of stealing the intellectual property of US AI labs to build Kimi K3 through a process known as "distillation," in which the outputs of one AI model are used as training inputs for another model. The terms and conditions of most AI vendors prohibit distillation, but legal experts say the tactic is a legal gray area rather than a clear case of IP theft.

Michael Kratsios, director of the White House Office of Science and Technology Policy, specifically said Moonshot had used outputs from Anthropic's Fable 5 to train Kimi K3.

Chinese Models in the OpenAI Incident

The strengths of Chinese AI models were also highlighted by the OpenAI rogue AI incident. Hugging Face, the company that came under attack, said it was forced to use a Chinese open source model — Z.ai's GLM-5.2 — to defend itself. Hugging Face said it first tried an unnamed American AI model but found its strict guardrails around cyber capabilities rendered it useless for conducting a defense against the ongoing attack.

When asked whether he found it concerning that Hugging Face had to use a Chinese-built model, Brockman did not directly answer the question. He reiterated the importance of having access to capable AI models for defense purposes.