OpenAI kicked off training its most advanced AI models, claiming it has strengthened security measures that are designed to prevent them from escaping restricted digital environments. The company reveals it has blocked live internet access during frontier model training to reduce the risk of AI agents reaching external systems without permission. Notably, the move comes after the AI firm found on September 20 that advanced AI agents could exploit software vulnerabilities and misconfigurations to bypass existing restrictions. Soon, OpenAI announced it was pausing training and evaluation of its most capable models to address these concerns.
Now, the company has once again resumed reinforcement learning training and internal model inference since October 1, after reportedly putting additional safeguards in place. In a new blog post, OpenAI explained why it is taking this approach. “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 builds an offline world for AI agents
AI agents need digital workspaces to function. But since giving AI models access to the live internet can pose security risks, AI farms developed offline environments called Sandboxes. That is, the offline digital world designed to keep the AI agents and their activities – writing code, searching for information and practising using software, etc. – separate from the active online world. But as AI models become more capable, they can also become better at finding weaknesses in the very systems meant to contain them.
To address this, OpenAI has built an offline world where models can interact with indexed and cached web content without freely accessing the live internet. This allows them to practise browsing and searching while reducing the risk of affecting external systems.
The company is also using multiple layers of protection, including restricted containers, virtual machines, network controls and monitoring systems. The idea is: even if an AI agent manages to bypass one security measure, other layers should stop it from going further.
AI agents found ways to bypass security controls
OpenAI also asked teams of advanced AI agents to test its defences in controlled environments. Their tasks included finding ways to access the live internet, cross security boundaries and communicate with agents they were supposed to remain isolated from.
The tests uncovered previously unknown software vulnerabilities and configuration errors. In one experiment, agents used processor cache timing differences to exchange information between isolated environments without opening a network connection or sharing writable files.
OpenAI says it has fixed the identified weaknesses, tightened network restrictions and improved systems that monitor suspicious activity. It also worked with open-source software maintainers to address vulnerabilities in affected components.
However, the company acknowledged that the work is far from over. As future AI models become more capable, they could discover weaknesses that remain unknown today. OpenAI says it will continue testing its defences before allowing its most advanced models to undergo further training and evaluation.
– Ends
Leave a Reply