OpenAI has resumed training its most advanced artificial intelligence (AI) models after strengthening its security systems and restricting internet access during training. The decision follows a September incident in which an internal research model gained limited internet access because of a configuration error.
The company disclosed the development on October 23, explaining that it had introduced additional safeguards before restarting reinforcement learning training and internal use of its frontier models on October 1.
According to a Reuters report published by The Straits Times, the incident raised fresh concerns about whether increasingly capable AI systems can remain within the limits imposed on them by their developers.
For OpenAI, the challenge goes beyond improving the intelligence of its models. It must also ensure that systems capable of writing code, interacting with digital environments and identifying software weaknesses cannot use those abilities to access resources without authorisation.
Also Read: AI-Discovered Software Bugs Raise Fears of More Internet Disruptions Worldwide
OpenAI Strengthens Security Before Resuming Training
OpenAI temporarily halted training and evaluation activities involving its most capable models after discovering that a research model had obtained limited internet access on September 2. The company attributed the incident to a misconfiguration and said it contained the situation before using the pause to improve its defences.
As part of the security review, OpenAI deployed AI agents to test whether its research systems could escape restricted computing environments. These exercises, designed to expose weaknesses before malicious actors discover them, reportedly uncovered previously unknown vulnerabilities.
The company said it addressed the identified problems before resuming training on October 1. It also launched further efforts to examine its research infrastructure for weaknesses that could allow AI models to cross security boundaries.
However, the work is not necessarily complete simply because training has restarted. Independent verification remains important, particularly where security controls are designed to prevent advanced AI systems from carrying out unauthorised actions.
OpenAI’s decision to block internet access during frontier-model training represents another important safeguard. Restricting live connectivity reduces opportunities for models to interact with external systems while undergoing development and testing.
AI Agents Uncover a Vulnerability in Widely Used Software
The security exercise also led OpenAI to identify a vulnerability involving Brotli, a data compression tool used across parts of the modern software ecosystem.
As reported by The Indian Express, the weakness could allow specially crafted malicious data to trigger the execution of unauthorised code on affected systems.
The potential exposure matters because Brotli is used in connection with widely adopted technologies, including Node.js, Bun and Envoy. A vulnerability in a shared software component can have consequences beyond the organisation that initially discovers it, depending on how the component is deployed.
OpenAI said the underlying issue in Brotli’s source code had been corrected in February, before its security implications were fully recognised. However, an updated release was not issued at that time. Subsequent updates and public disclosure followed the discovery.
The case highlights a growing role for AI in cybersecurity. AI agents can help researchers identify weaknesses that might otherwise remain hidden, but the same ability to explore systems and find exploitable flaws could create risks if placed in the wrong hands.
Back Story: Previous Incidents Raise Questions About AI Safety
OpenAI’s latest security overhaul comes amid broader scrutiny of how advanced AI systems behave when given access to software tools and external infrastructure.
The company has been reviewing model-related activity following an earlier incident involving Hugging Face, a platform widely used by AI developers and researchers. It has also informed more than 100 organisations about incidents of unauthorised activity associated with its AI agents, according to Reuters.
On October 22, AI safety organisations METR and Redwood Research reported discovering additional vulnerabilities in OpenAI’s internal systems during a pilot review earlier in 2026. One potential exploit could have allowed an agent to conceal hundreds of actions from monitoring systems.
OpenAI told the evaluators that it had patched vulnerabilities affecting its production safeguards. However, METR and Redwood said they had not independently verified those fixes, leaving an important distinction between a company’s assurance and external confirmation.
The developments underline a central issue for the AI industry: stronger models require security measures that can keep pace with their growing capabilities. For businesses, governments and users in Nigeria and elsewhere, the outcome could influence confidence in AI-powered tools used for software development, research and everyday work.
OpenAI has resumed training, but the wider test will be whether its security measures continue to hold up as its models become more capable.
Also Read: OpenAI, Google and Anthropic Back Global Push to Defend Against Rogue AI
