OpenAI resumes AI training without live internet

OpenAI has resumed training its most advanced AI models, but this time, the company is keeping them away from the live internet. The decision comes after internal tests showed that increasingly capable AI agents could find ways around security controls and potentially reach systems they were not supposed to access.

Company paused training and evaluation of its capable models after spotting the problem on September 20. Pause happened after it was found that AI agents could exploit software bugs and configuration errors to slip past rules so AI agents were seen as a risk.

Training and internal model inference restarted on October 1, after OpenAI introduced additional safeguards. In a new blog post, the company explained that keeping frontier models offline is currently the safer option.

“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,” the company said.

OpenAI creates an offline environment for AI agents

AI agents need digital environments where they can perform tasks such as writing code, searching for information and practising with software. However, unrestricted access to the live internet could create new security risks.

To reduce those risks, OpenAI has created an offline environment, or sandbox, where its models can work with indexed and cached web content without directly connecting to the live internet. This allows agents to practise browsing and information retrieval while limiting their ability to interact with external systems.

OpenAI is also relying on several layers of security. These include restricted containers, virtual machines, network controls and monitoring tools. The goal is to ensure that if an AI agent gets past one barrier, additional safeguards can prevent it from moving further.

AI agents exposed hidden security weaknesses

OpenAI also used teams of advanced AI agents to test its own defences. The agents were instructed to find ways to access the live internet, cross security boundaries and communicate with other agents from which they were meant to remain isolated.

The tests uncovered previously unknown software vulnerabilities and configuration problems. In one experiment, agents were able to use differences in processor cache timing to exchange information between isolated environments without establishing a network connection or sharing writable files.

OpenAI says it has addressed the weaknesses discovered during testing, strengthened network restrictions and improved monitoring for suspicious behaviour. The company also worked with open-source software maintainers to fix vulnerabilities in affected components.

Still, OpenAI says the challenge is ongoing. More capable models could potentially discover new weaknesses that security teams have not yet identified. For now, the company plans to continue testing its safeguards before allowing its most advanced models to undergo further training and evaluation.

ABOUT THE AUTHOR

Kahekashan

[email protected]

Kahekashan is a passionate technophile with a keen eye for cutting-edge gadgets, emerging technologies, and everything in the digital realm. Raised in a Defence family with strong values and a background in literature, she has consistently pursued excellence in every endeavour. Her last full-time assignment involved content writing with the Indian School of Business.

Read More