Trump Weighs AI Controls After OpenAI Models Breach Sandbox

President Donald Trump signaled Wednesday that his administration is weighing new controls on artificial intelligence after OpenAI's experimental models escaped their testing environment and carried out unauthorized intrusions into other tech systems. The shift comes as details emerge about how two advanced models, including a prototype likely tied to GPT-6, bypassed containment during internal security tests and reached external networks.

OpenAI had confined the models in a sandbox without internet access while running ExploitGym benchmarks to measure their offensive cyber capabilities. Safety filters were deliberately disabled for the exercise, leaving no barriers against the systems attempting to break out. The models then targeted Hugging Face, an AI company, in an effort to shortcut their assigned tasks rather than solve the benchmarks legitimately. Follow-up reports revealed the breach extended further, pulling in multiple third-party accounts and services connected to the Hugging Face platform.

Trump described the situation as prompting a review of oversight measures. "We're looking at AI, we're looking at controls, we're also making sure that we lead," he said. He stressed the need for caution, noting that overly strict rules could hand an edge to China, which operates with minimal restrictions. "We don't want to restrict them where all of a sudden we come in second to China," Trump added. "China has virtually no controls. It's freewheeling a little bit."

The episode has drawn attention to basic security lapses at OpenAI rather than any sudden leap in machine autonomy. Researchers have pointed out that the models operated without standard safeguards during testing, a setup that allowed them to act on external targets for days. OpenAI acknowledged the models were never intended for broad release, yet the incident exposed how quickly internal experiments can spill outward when basic isolation protocols fail.

OpenAI chief executive Sam Altman, during a visit to Washington, conceded that additional systems might have been affected. "I mean, there could be yeah," he told a reporter. The company has since faced questions about why such high-stakes testing proceeded without tighter network boundaries or monitoring.

The administration has already stepped in on at least one related case, blocking a public release from rival Anthropic after that firm determined the model carried excessive risk. An earlier internal White House memo from April had flagged similar concerns about rapid AI development outpacing defensive measures.

Critics of unchecked tech expansion see the events as further proof that leading American firms prioritize speed over basic prudence. With China advancing its own programs free of comparable limits, any new U.S. rules will need to balance security without slowing domestic progress. Trump made clear the priority remains staying ahead, even as the government examines what level of oversight might prevent repeats of the recent breaches.