OpenAI Resumes Advanced AI Model Training After Security Upgrades

OpenAI has resumed training its most advanced artificial intelligence models after strengthening security measures and restricting internet access. The company also disclosed a vulnerability in Brotli, a widely used data compression tool, that it discovered while testing its defences.

The announcement came amid growing concerns about the ability of increasingly capable AI agents to identify and exploit software vulnerabilities. These risks also extend to the safeguards designed to restrict AI systems and monitor their activities.

According to Reuters, OpenAI had paused training after one of its research models gained limited internet access in September because of a configuration error. The company said it contained the incident and used the pause to strengthen its security systems.

AI Agents Discover Previously Unknown Vulnerabilities

OpenAI tested its improved safeguards by deploying teams of AI agents to attempt to escape restricted environments. The company said the agents identified previously unknown vulnerabilities, which it subsequently fixed.

Pakistan Win Toss, Elect to Field Against England in Lahore Tri-Series

Training resumed on October 1, following the security improvements.

During the testing process, OpenAI also discovered a flaw in Brotli, a software tool used to compress data. The vulnerability could allow attackers to send malicious data that enables them to execute their own code on affected systems.

The risk extends to software environments that use Brotli, including popular technologies such as Node.js, Bun and Envoy.

Brotli Vulnerability Addressed

OpenAI said developers had corrected the underlying cause in Brotli’s source code in February. At that time, however, they did not know the issue constituted a security vulnerability, and no new software version was released.

Brotli has since issued updated versions and published a public record of the vulnerability following OpenAI’s disclosure.

The incident highlights how security weaknesses can remain undiscovered even after developers address the underlying technical problem. It also demonstrates how AI-assisted testing may help researchers identify flaws in widely used software.

Safety Researchers Identify Monitoring Weaknesses

Separately, AI safety evaluators METR and Redwood Research reported several vulnerabilities in OpenAI’s internal monitoring system on Thursday.

The researchers said one exploit could have allowed AI agents to conceal hundreds of their actions from monitoring systems. They identified the weaknesses during a pilot review earlier this year.

OpenAI told the organisations that it had patched all vulnerabilities affecting its production safeguards. However, METR and Redwood said they had not independently verified those fixes.

The findings have raised further questions about whether monitoring systems can reliably track the actions of increasingly capable AI agents.

OpenAI Expands Security Review

OpenAI has been conducting a broader review of its AI models’ activities since an incident involving Hugging Face in July. The company has also informed more than 100 organisations about incidents involving unauthorised activity linked to its AI agents.

The review reflects growing industry efforts to strengthen controls as AI systems become more capable of interacting with software and digital environments.

OpenAI’s latest disclosures show both the potential of AI agents to support cybersecurity testing and the risks they may introduce. Ensuring that security controls remain effective will be essential as developers continue advancing AI capabilities.