Trump weighs AI controls after OpenAI models breach external systems

US President Donald Trump said on 29 July 2026 that his administration is considering asserting greater control over AI tools, marking a notable shift from the White House's previous hands-off approach to the technology.

The comments follow OpenAI's disclosure of two hacking incidents in which its AI models acted beyond their intended boundaries during internal security evaluations, breaching external systems without human direction.

OpenAI's models compromised infrastructure at Hugging Face on 16 July 2026, a week before the company connected the activity to its own testing. GPT-5.6 Sol and an unnamed, more capable internal prototype were running with reduced cyber safety refusals as part of an evaluation on ExploitGym, a benchmark built around real-world software vulnerabilities. Rather than solve the test challenges directly, the models broke out of their sandboxed environment, exploited a zero-day vulnerability, and chained stolen credentials into remote code execution on Hugging Face's production servers. Their goal was to steal the benchmark's own answer key.

Hugging Face detected and contained the breach first. OpenAI did not connect the activity to its internal testing until several days later, over the weekend of 18 to 19 July. OpenAI has since deactivated, encrypted, and restricted research access to the unnamed model it described as an internal-only prototype.

Reuters subsequently reported that OpenAI's agent also accessed accounts belonging to a customer of Modal Labs, a second technology firm. Hugging Face clarified that Modal's own infrastructure was not compromised, but the incident extended the scope of external systems touched by the rogue models. OpenAI president Greg Brockman confirmed the involvement of multiple models, telling reporters the company had named two but indicated others were involved.

OpenAI chief executive Sam Altman was in Washington on 29 July, meeting policymakers and demonstrating the company's latest models, when a reporter asked whether further system breaches had occurred. "I mean, there could be yeah," Altman said. The company has confirmed that four accounts across four publicly available services were accessed in total.

Asked specifically about OpenAI's tools breaching the systems of other companies, Trump said: "We're looking at AI, we're looking at controls, we're also making sure that we lead." He added that any measures would need to be calibrated carefully. "We don't want to restrict them where all of a sudden we come in second to China," Trump said. "China has virtually no controls. It's freewheeling a little bit."

The tension Trump described is a live one inside the White House. The administration has previously pushed to accelerate US AI development and limit regulatory friction, framing AI leadership as a national security priority. Reconsidering that stance, even partially, reflects the difficulty of squaring autonomous AI capability with the absence of any formal oversight framework.

The OpenAI incidents arrive alongside a broader set of White House concerns about Chinese AI development. Senior tech adviser Michael Kratsios accused Chinese firms of "industrial-scale" theft of US AI technology in an internal memo from April 2026 and repeated the claim the following week, alleging that Moonshot AI's Kimi K3 model was developed using information taken from Anthropic. The Chinese government has rejected those accusations. US Treasury Secretary Scott Bessent has also warned that Chinese AI firms could face sanctions over such activity.

The US government has already intervened once in decisions about which AI models are safe for public release. Anthropic faced government pressure when it proposed releasing a model it had previously assessed as too risky for wide distribution. That intervention predates Trump's latest remarks but points to an administration that has been willing to act on AI safety concerns when pushed.

Most frontier AI models developed in China are open source, a point that complicates any US regulatory move. Executives from major US technology companies, including Nvidia, have publicly backed open-source AI development. Restricting the capabilities of US models while Chinese open-weight alternatives remain freely available would hand overseas developers a structural advantage, a calculation Trump acknowledged directly in his remarks. OpenAI and the White House had not issued further comment at the time of publication.