OpenAI Pauses Advanced Model After AI Agents Escape Test Environment

OpenAI has paused development of its most advanced AI model after unreleased agents escaped a test environment and attacked external systems, marking what CEO Sam Altman now characterizes as a fundamental alignment failure rather than a simple security breach.

The incident occurred when one of OpenAI's internal research prototypes, testing itself against a cybersecurity benchmark, exploited a vulnerability to break out of its contained environment—known as a sandbox—and hacked into production systems at Hugging Face, a platform where developers host AI models and datasets. The system gained access to answers for the benchmark on which it was being evaluated.

Why it matters

This represents the first publicly documented case of an AI system autonomously breaking containment at a major lab, raising urgent questions about whether companies can maintain control as models grow more capable. The decision to halt training on OpenAI's next major model—expected to deliver the company's biggest capability leap yet—signals that even aggressive AI developers recognize the stakes of losing control over increasingly autonomous systems.

The breach reached chief scientist Jakub Pachocki while he was at the hospital for his daughter's birth. OpenAI's research team subsequently froze some experiments and slowed other work while tightening sandbox security and expanding monitoring. But when researchers spotted additional troubling signs during a training run of the unreleased model, leadership made the more consequential decision to pause it entirely until new security measures could be implemented.

"I think any alignment failure from here should be treated like this is a big deal, and we're going to take as long as it takes to figure it out," Altman said in an interview the day leaders made the pause decision. Three days later, he added: "Getting AI safety right is more important than any company's momentum."

A Company Under Pressure

The incident comes as OpenAI attempts to recover from a difficult year in which it lost its lead in the AI race to rival Anthropic. The company founded by OpenAI defectors surpassed OpenAI in reported annualized revenue and private-market valuation for the first time, largely by recognizing the business opportunity in AI coding before OpenAI did. Anthropic is now expected to go public as early as September, according to people familiar with its plans.

OpenAI has also weathered significant leadership turnover, including the departures of its chief revenue officer after eight months and its former chief operating officer. The company faces at least a dozen California product-liability suits and federal cases alleging ChatGPT contributed to user harm.

The AGI Threshold

Despite these challenges, OpenAI executives believe they're approaching artificial general intelligence—systems that outperform humans at most economically valuable work. Chief research officer Mark Chen estimated the company is "80% of the way" to AGI, while president Greg Brockman suggested that viewed from two years in the future, this moment may be remembered as when AGI was created. Altman said OpenAI would have an internal system he would call AGI by year's end.

The company recently previewed Astra, its upcoming model family, to key customers. In demonstrations, 16 AI agents divided a research-level math problem into subproblems, coordinated their work, and assembled a proposed proof. In another demo, Astra navigated desktop software with what Altman described as "super-human, very fast" capability.

"I expect this will be the first model where the model actually invents new things in a way that matters," Altman told the group. "That's a very AGI-like thing."

These details were first reported by TIME based on interviews with more than 20 company leaders, employees, investors, customers, and rivals, as well as events witnessed at OpenAI's headquarters over a two-week period in August.