OpenAI resumes frontier AI training after security overhaul, discloses flaw in widely used software

OpenAI said on Friday that it had restarted training its most advanced AI models after improving security and blocking internet access. The company also revealed a flaw in widely used software that it found while testing its defenses.

The news comes as people worry about the ability of increasingly powerful AI agents to find and use software flaws, including weaknesses in the safety measures designed to keep them under control.

OpenAI had stopped training its most advanced models after one of its research models gained limited internet access in September because of a configuration mistake. The company said it contained the incident and used the pause to strengthen its defenses.

To test these defenses, OpenAI had teams of AI agents try to escape from the restricted environments they were placed in. The company said the agents found flaws that had not been known before. OpenAI fixed these flaws, and training resumed on October 1.

During this test, the company behind ChatGPT found a flaw in Brotli, a widely used data compression tool.

OpenAI said attackers could use the flaw by sending harmful data that allows them to run their own code on affected systems. These systems include those that use popular web software such as Node.js, Bun and Envoy.

OpenAI said the main cause of the flaw was fixed in Brotli's source code in February, before anyone knew it could create a security risk. However, no new software version was released at that time. Brotli has since released updated versions and made a public record of the flaw after OpenAI disclosed it.

Separately, on Thursday, AI safety groups METR and Redwood Research said they had found several new weaknesses in OpenAI's internal system. These included a way AI agents could have used to hide hundreds of their actions from monitoring during a trial review earlier this year.

OpenAI told the safety groups that it had since fixed all weaknesses affecting its systems used to protect its live services. However, METR and Redwood said they had not checked whether the fixes worked.

OpenAI has been carrying out a broad review of activity by its AI models since the Hugging Face incident in July. It has also informed more than 100 organizations about incidents involving unauthorized activity by its AI agents.