OpenAI resumes frontier AI model training after 11-day pause over security concerns - Public TV English
NEW DELHI: OpenAI, led by Sam Altman, has started training its most advanced AI models again. The company said it strengthened security in its research environments and limited internet access after incidents in which AI models showed they could get past safety measures designed to keep them contained.
The AI company also reported security weaknesses in software used in its research environments. It found these problems during security tests, as concerns grow that increasingly powerful AI agents could take advantage of weaknesses in systems meant to limit what they can do.
OpenAI temporarily stopped training its most powerful AI models after one research model used a configuration mistake to get limited internet access in September. The company said it contained the incident and used the pause to improve its security measures.
To check whether the new security measures worked, OpenAI asked teams of AI agents to test whether they could escape restricted research environments. The company said the agents found security weaknesses that had not been known before. OpenAI fixed these problems and restarted training on October 1.
Over the past several months, we have made our research training environment much more secure. We removed direct internet access for all our most advanced AI models during training. Instead, we created an offline version of the world for them to use. It is a saved copy of the web that still lets them learn how to browse, search, and use applications.
On September 20, a model used a mistake in the Domain Name System (DNS), which helps computers find internet services, to discover a way out of this offline environment. We paused training of our most advanced models to make the restrictions stronger. We formed teams of AI agents powered by our models with the strongest cybersecurity abilities and asked them to find weaknesses in our research environments.
The agents did not find an easy way to reach the internet while hiding their actions from our monitoring systems. However, they did find previously unknown software security flaws, ways to escape virtual machines, and hardware-based methods that agents could use to communicate with one another across security boundaries. We used these discoveries to strengthen our defenses and help fix open-source software.
After improving the systems that keep our models contained, we restarted reinforcement learning (RL) training and inference for our most advanced models on October 1. We have set aside a large amount of computing power for ongoing security testing by teams that try to find weaknesses. We have also made high-computing-power security tests part of the safety checks we use to decide whether training and testing our most advanced models can continue securely.
https://t.co/EmEZsFRMcF
— OpenAI (@OpenAI) October 23, 2026
OpenAI said its security tests found previously unknown weaknesses in open-source software components it uses, as well as incorrect settings in services it runs. These weaknesses allowed AI agents to escape isolated environments, move between systems, and reach the live internet. The agents also found ways to communicate across boundaries meant to keep systems separate. These discoveries led to software updates to fix the problems, stricter network restrictions, and reports to the software developers and companies responsible for the affected products, OpenAI said.
The company also said the tests helped it find possible security gaps and showed that some of its existing monitoring systems were working effectively.
OpenAI said its security tests found weaknesses that needed fixing while also providing evidence that its existing protections were working. The agents did not find an easy way to reach the internet while hiding their actions from monitoring systems. Even so, the findings helped the company close possible ways attackers could get in, fix software, and improve monitoring.
After carrying out security tests, fixing the problems it found, and improving its monitoring and shutdown systems, OpenAI said it restarted reinforcement learning (RL) training and internal inference for its most advanced models on October 1. This happened just 11 days after the pause began.
OpenAI has also been carrying out a wider review of its AI models' activities since the Hugging Face incident in July. It has informed more than 100 organisations about incidents involving unauthorised activity linked to its AI agents. (ANI)