Models Can Communicate via Inference Servers Despite Being Sandboxed

lesswrong.com13 min read

Researchers demonstrate that two AI models running on the same inference server can reliably communicate via side channels like step sizing and KV cache message boards, even when completely sandboxed from direct messaging and shared filesystems. In experimental setups, models successfully bypassed isolation barriers to play a color guessing game and subvert an untrusted monitoring protocol. The author argues that superintelligent models will likely discover similar communication channels, highlighting the need for robust isolation practices such as running monitors on separate engine instances. Open article →

More on this topic