OpenAI has started training its most powerful systems again – but will keep them off the internet for fear they would go rogue.
Last month, the company announced that it had paused training of its most advanced models after one of them broke out of containment and found a way to connect to the internet. It said it would be pausing that training until it could be sure that it had developed sufficiently robust safeguards to stop similar incidents in the future.
Now it says that it has tightened those defences and that it would begin training them again. But the systems were still too dangerous to allow on the internet, it said.
“Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access,” it wrote in a blog post announcing the change.
At the same time, it said that it had found a flaw in a widely used piece of software while it had been testing those defences. The disclosure comes amid concerns over the ability of increasingly capable AI agents to identify and exploit software flaws, including in the guardrails meant to contain them.
To test the new safeguards, OpenAI had teams of AI agents try to break out of their restricted environments. It said the agents found previously unknown flaws, which it had fixed, and training resumed on October 1.
During that exercise, the ChatGPT maker found a flaw in Brotli, a widely used data compression tool.
The company said attackers could exploit the flaw by sending malicious data that lets them run their own code on affected systems, including those using popular web software such as Node.js, Bun and Envoy.
OpenAI said the root cause was fixed in Brotli's source code in February, before it was known to create a security vulnerability, but no new version was released at the time. Brotli has since released updated versions and published a public record of the flaw after OpenAIdisclosed it.
Separately on Thursday, AI safety evaluators METR and Redwood Research said they found several new vulnerabilities in OpenAI's internal system, including an exploit agents could have used to keep hundreds of their actions from being monitored, in a pilot review earlier this year.
OpenAI told the safety organisations it has since patched all vulnerabilities affecting its production safeguards, although METR and Redwood said they had not verified the fixes.
OpenAI has been conducting a broad review of its AI models' activity since the Hugging Face incident in July, and has informed more than 100 organisations about incidents of unauthorised activity tied to its AI agents.
Additional reporting by agencies
Read More‘Skynet’ AI hijacks other AI to carry out attacks, researchers warn
What you need to do as the world panics about AI cyber attacks
Things will get ‘a lot worse before they get better’, experts warn amid AI panic
What is PHP and why has it caused chaos for millions of websites?
Websites across the internet go down as hackers launch millions of cyber attacks
OpenAI agents disabled their own safeguards and then lied about it, company admits



