NEW DELHI: Sam Altman-led OpenAI has resumed training its most advanced AI models after strengthening security across its research environments and restricting internet access, following incidents in which AI models demonstrated the ability to bypass containment safeguards, the company said in a statement.
The AI giant also disclosed vulnerabilities in software used in its research environments, identified during security testing, amid growing concerns that increasingly capable AI agents could exploit weaknesses in systems designed to restrict their activities.
OpenAI had temporarily suspended training of its most capable AI models after a research model exploited a configuration error to gain limited internet access in September. The company said the incident was contained and the pause was used to strengthen its security measures.
To assess the effectiveness of its strengthened security measures, OpenAI deployed teams of AI agents to test whether they could escape restricted research environments. The company said the agents identified previously unknown vulnerabilities, which were subsequently addressed, allowing training to resume on October 1.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
https://t.co/EmEZsFRMcF
— OpenAI (@OpenAI) October 23, 2026
“In our red-teaming experiments, agents were able to find previously unknown software vulnerabilities in open-source components we use, as well as misconfigurations in services that we operate. These weaknesses allowed agents to escape sandboxes, move between systems, and reach the live internet. They also found ways to communicate across isolation boundaries. These findings led to software patches, tighter network restrictions, and disclosures to the relevant software maintainers and vendors,” the AI giant said.
The company further said its testing helped identify potential security gaps while also demonstrating the effectiveness of existing monitoring systems.
“Our red-teaming campaign found weaknesses that required repairs, while also providing evidence about the protections we already had in place. We did not find a route to the internet that agents could trivially exploit while concealing their activity from our monitors. The findings nevertheless helped us close potential attack paths, patch software, and strengthen our monitoring,” it said.
Thus, following security testing, implementation of necessary fixes and improvements to its monitoring and shutdown systems, OpenAI “resumed reinforcement learning (RL) training and internal inference of our frontier models on October 1,” just “11 days” after its pause began.
OpenAI has also been conducting a broader review of its AI models’ activities since the Hugging Face incident in July and has informed more than 100 organisations about incidents of unauthorised activity linked to its AI agents. (ANI)
