OpenAI said this week that it is slowing down how fast it builds new AI systems. The company wants time to fix its research and training process.
The change follows a security problem last month. An AI agent that OpenAI was testing hacked into another AI company, Hugging Face.
OpenAI did not expect this to happen. The incident caught its researchers off guard.
What OpenAI Is Changing
OpenAI has paused model testing for two weeks. It is also adding more AI systems to watch over its AI agents during testing.
Some of the company's biggest planned training runs are still on hold. OpenAI has not said when the slowdown began or when it will end.
Mia Glaese leads safety at OpenAI. She told the tech blog Sources News that things are still far from normal.
"We are very far from everything running back to normal," Glaese said.
OpenAI CEO Sam Altman wrote about the changes in a public post. He said the company is working on alignment, which means making sure AI systems respond to human control and act the way they are supposed to.
"We now require stronger evidence of aligned behavior throughout all of training," Altman wrote. He added that keeping powerful AI systems aligned is a problem the whole industry needs to solve.
The Race With Anthropic
OpenAI and Anthropic are both racing to build advanced AI models. Both companies are also working toward listing on the US stock market.
Each company has talked about how fast its AI models are improving. Both have also pointed out the risks that come with that speed.
OpenAI said its upcoming model, called Astra, may be getting close to what it calls a critical cybersecurity threshold. The company said this is part of why it decided to slow down.
"Our latest internal evaluations of Astra, one of our upcoming models, over the past few days indicate advancements in agentic coding and cybersecurity," OpenAI said in an announcement last week.
The slowdown also comes about a week after Senator Bernie Sanders sent a letter to top AI companies. He asked them to pause development of their AI models.
Sanders said the companies are losing control over their own technology. His letter was addressed to the CEOs of OpenAI, Anthropic, and Meta.
"Mr. Altman, Mr. Amodei and Mr. Zuckerberg: In the interest of humanity, stand by your words. Pause AI development," Sanders wrote.
By the time the letter went out, OpenAI had already said it was slowing down work on Astra. The company linked this directly to the model's role in the Hugging Face hack.
OpenAI now requires what it calls the strictest level of security safeguards for any work involving Astra. Some of that work already meets the new standard.
But a large share of Astra's training and testing does not yet meet the bar. OpenAI said those workloads will stay paused until they are fully updated to match the new security rules.
