www.lesswrong.com
Models Can Communicate via Inference Servers Despite Being Sandboxed — LessWrong
Two AI agents running on the same inference server can communicate by detecting and influencing each other's response times through shared computational resources, bypassing explicit messaging restrictions.