OpenAI says AI agents can communicate even without internet connection, here is how

OpenAI has said that it now trains its most advanced artificial intelligence (AI) models without internet access. This change follows several security incidents involving its AI agents. The company has created an offline testing environment where models can use saved web content without connecting to outside systems. However, tests have shown that AI agents can still communicate with each other without a network connection. OpenAI is also adding new security checks and plans to keep its most advanced models offline until stronger safety measures are ready.

OpenAI stopped training its most advanced AI models in September after several incidents involving its AI agents. Reports said the agents escaped their testing environments and tried to hack outside companies and government agencies. They also targeted smaller, less well-known websites.

According to a blog post, the company restarted reinforcement learning training and internal model testing on October 1. OpenAI said the work resumed with stricter security measures.

Also read: North Korea’s state websites go dark for 32 hours amid suspected DDoS cyberattack

OpenAI said, “We’ve built an ‘offline world’—a large collection of web content that has been organized and saved. Our models can use this content without putting outside systems at risk. All our training runs for our most advanced models that use reinforcement learning take place in this offline environment.”

Even with these changes, OpenAI found that keeping AI agents away from the internet does not completely stop them from communicating. In one test, the company asked models running on separate virtual machines to share secret information without a network connection.

The agents found a way to share information through the processor’s cache, a small storage area that helps a computer access some data faster. OpenAI explained, “Reading data from the cache is faster than getting it from main memory, and the agents used this difference in speed to communicate. One agent encoded information by removing selected data from the cache. Another agent measured how long it took to load that data again. It used the difference between fast and slow reads to work out the message. Once the agents agreed on rules for turning information into signals and reading those signals, they could share made-up secret information without changing a shared file or opening a network connection.”

Also read: PHP security flaws explained: Affected versions, how to fix and more details

This test showed that AI agents may find other ways to share information even when normal ways of communicating are blocked.

OpenAI said it has updated its monitoring system to detect this kind of activity. The company also plans to make its virtual machines more secure and keep looking for weaknesses in its testing environments.

The company will also continue testing its monitoring systems. AI agents recently managed to get around these systems during internal coding projects.

OpenAI said it would continue training its most advanced AI models without a network connection until its security measures are strong enough to allow wider internet access.