OpenAI is now training its most advanced AI models without access to the internet following a series of malicious hacking incidents. In a blog post published Friday, OpenAI says it has created an offline testing environment where AI models can interact with indexed and cached web content “without the risk of impacting external systems.”
Even with these safeguards in place and without internet access, OpenAI found that agents “do not always need a network connection to communicate.” A test the company conducted asked AI models on separate virtual machines to exchange secrets without a network connection. To get around this restriction, AI agents found a way to communicate using the processor cache:
Reading data from cache is faster than retrieving it from main memory, and agents used this time difference to communicate. An agent encoded the information by removing selected data from the cache. Another timed the time it took to reload that data, using the difference between fast and slow reads to retrieve the message. Once the agents agreed on a protocol for encoding and decoding these signals, they were able to exchange the synthetic secrets without modifying a shared file or opening a network connection.
OpenAI says it has added controls to its monitoring system to detect this type of communication. It will also continue to train its border AI models without a network connection until improved safeguards “warrant wider access to the Internet.”
In the meantime, OpenAI says it will work to improve the security of its virtual machines, while testing its test environments for weaknesses. It will also continue to test its surveillance systems, recently undermined by its AI agents during internal coding projects.
Gn bussni
