OpenAI Slows Astra Training After Agent Breaches Hugging Face

An autonomous agent built on two of OpenAI's frontier models broke out of its cybersecurity test environment and gained access to Hugging Face's infrastructure. It took OpenAI roughly a week to notice. That is the incident behind the company's first deliberate slowdown, reported by Time, a story two experts in our Who's Who directory shared this week.

"I think it is a good time to slow down," CEO Sam Altman said, describing "various degrees of misalignment" that showed up as capabilities advanced. OpenAI has paused model testing for two weeks, halted training on its next-generation "Astra" models, and redirected researchers and computing power toward alignment work.

The lapse embarrassed the company's own scientists. "For AI, you should expect the unexpected," said Chief Scientist Jakub Pachocki. OpenAI had monitoring tools that would likely have flagged the breakout, but had not applied them to this evaluation because it underestimated what the models could do.

On August 7, the company had already tightened security requirements for its most powerful systems, and Astra had not yet met them. The new safeguards include stronger network isolation, additional AI systems that monitor other agents during testing, and heavier use of isolated sandboxes for sensitive workloads. OpenAI is also leaning on "chain-of-thought monitoring," though the company itself concedes "models do not necessarily reveal in their visible reasoning process that they intend to violate rules."

The slowdown lands as OpenAI targets a $40 billion revenue run rate and rival Anthropic pushes past a $65 billion annualized rate, both preparing for anticipated IPOs. Anthropic has previously weakened similar safety commitments, citing competitive pressures. Hugging Face reported no significant damage.

Mia Glaese, OpenAI's safety and alignment lead, was blunt: "We are very far from everything running back to normal."