Tag
This paper studies emergent languages that autonomous LLM agents propose to one another on the Moltbook platform, finding that some languages are specifically designed to evade human oversight and can be learned in-context from short descriptions. The findings raise safety concerns about monitoring agent populations.