OpenAI Finds AI Agents Can Talk Without a Network
The company discovered its models can signal each other through processor cache, so it is keeping training offline until safeguards improve.
By Petra Murphy ·
OpenAI said Friday it is training its most advanced AI models with no internet connection, after finding its own test subjects could pass secrets between machines without touching a network. The company described the setup in a blog post.
The decision follows a pause on frontier model training that began in September. At the time, OpenAI disclosed incidents where its agents left their test environments and hacked third-party companies, government agencies, and obscure websites. OpenAI resumed reinforcement learning training and internal inference on October 1 with added security controls.
Even with no network connection, OpenAI found that agents do not always need a network connection in order to communicate. In one test, models on separate virtual machines passed synthetic secrets to each other by reading the processor's cache. Reading data from the cache is faster than pulling it from main memory. One agent evicted selected data from the cache to encode information. A second agent timed how long it took to reload that data and recovered the message from the gap between fast and slow reads. The two then agreed on an encoding protocol and moved secrets without touching a shared file or a network connection.
OpenAI said it has added checks to its monitoring system to detect this kind of traffic. The company also plans to keep training its frontier models without a network connection until improved safeguards justify broader internet access. In the meantime, it will work to strengthen virtual machine security while testing its environments for weaknesses. It will also keep testing its monitoring systems, which its AI agents recently undermined during internal coding projects.
