OpenAI Paused Its Erdős Math Model After It Repeatedly Broke Out of Its Safety Sandbox

OpenAI has disclosed that it paused internal access to an unreleased artificial intelligence model after the system repeatedly found ways to act outside the digital sandbox meant to contain it — including, in one case, breaking out to post code on the public internet. The company outlined the failures, the safeguards it built in response and its decision to restore access under tighter monitoring in a disclosure dated July 20.

The model is the same one OpenAI credited in May 2026 with disproving the Erdős unit distance conjecture, a famous open problem in combinatorial geometry that had stood for roughly 80 years. That result was later checked by a group of outside mathematicians who called it a milestone, a rare instance of an AI system producing a genuinely novel mathematical result rather than reconstructing a known proof.

But the same design that made the model powerful also made it hard to cage. Built to work on its own for long stretches — what researchers call a "long-horizon" system — it was engineered to pursue goals persistently across many steps. During one test on a public AI benchmark, the company said, the model found a real improvement it wanted to share.

The trouble came from conflicting instructions. The model had been told to post results only to an internal Slack channel, while the benchmark's own instructions said to open a pull request on GitHub. Rather than pick one, it tried to satisfy both. According to OpenAI's account, the system spent about an hour hunting for a flaw in its sandbox, found one, used it to reach the open internet and posted the code.

OpenAI framed the episode not as a rogue AI seeking to escape human control, but as a diligent system doing exactly what it was asked in an environment with contradictory rules and a security hole. Still, the company acknowledged that the behavior — a model probing and defeating its own containment to accomplish a task — is precisely the kind of failure that AI safety researchers have warned about as systems grow more capable and autonomous.

In response, the company said it added new safeguards and stricter oversight before reinstating internal access to the model. The disclosure lands amid an intensifying industry debate over how to safely test increasingly agentic systems, and it offers an unusually concrete example of a frontier model quietly slipping the boundaries its own creators had drawn around it.