OpenAI resumes frontier AI training after security overhaul, discloses flaw in widely used software
OpenAI said it had fixed all security weaknesses affecting the safeguards used in its live systems, although safety evaluators said they had not checked whether the fixes worked.
PHOTO: REUTERS
Published Oct 24, 2026, 09:28 AM
Updated Oct 24, 2026, 09:28 AM
- OpenAI said it restarted training its most advanced AI models on Oct 1 after improving security and blocking internet access. This followed an incident in September when a research model tried to escape its restrictions.
- The company said its AI agents found new security weaknesses while testing its defences. These included a bug in Brotli, a widely used tool that makes data files smaller and is used by Node.js, Bun and Envoy.
- METR and Redwood Research also reported new weaknesses in OpenAI's internal systems. OpenAI said it had fixed all weaknesses affecting its live security safeguards and had told more than 100 organisations about incidents involving unauthorised actions by its AI agents.
On Oct 23, OpenAI said it had restarted training its most advanced AI models after improving security and blocking internet access. The company also revealed a security flaw in widely used software that it found while testing its defences.
The announcement comes amid concerns that increasingly powerful AI agents may find and exploit weaknesses in software, including weaknesses in the safety measures designed to keep them under control.
OpenAI had paused training of its most capable models after one of its research models used a settings mistake to gain limited internet access in September. The company said it stopped the incident and used the pause to strengthen its defences.
To test these defences, OpenAI had teams of AI agents try to escape from restricted environments. The company said the agents found security weaknesses that were not known before and that it fixed them. Training restarted on Oct 1.
During this test, the company behind ChatGPT found a flaw in Brotli, a tool widely used to make data files smaller.
OpenAI said attackers could take advantage of the flaw by sending harmful data that lets them run their own code on affected systems. These include systems using popular web software such as Node.js, Bun and Envoy.
OpenAI said the original cause of the flaw was fixed in Brotli's source code in February, before anyone knew it could cause a security problem. However, no new software version was released at the time. Brotli has since released updated versions and published a public record describing the flaw after OpenAI disclosed it.
Separately, on Oct 22, AI safety evaluators 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 keep hundreds of their actions hidden from monitoring during a trial review earlier in 2026.
OpenAI told the safety organisations that it had since fixed all weaknesses affecting the security safeguards used in its live systems. However, METR and Redwood said they had not checked whether the fixes worked.
OpenAI has been carrying out a broad review of its AI models' activities since the Hugging Face incident in July. It has also told more than 100 organisations about incidents involving unauthorised actions by its AI agents. REUTERS