Does Safety Molt? Evaluating LLM Safety in Multi-Agent Social Environments

Large language model safety is typically evaluated in isolation, yet deployed AI agents increasingly operate in persistent social environments where they interact with other agents over long horizons. Does a model that appears safe alone remain safe in a society of agents? We investigate this question through privacy by introducing a Moltbook style simulation platform in which 2,533 LLM agents interact across 124 communities over 25 simulated days, complemented by a controlled evaluation spanning seven frontier models and 7,000 evaluation traces. Our results reveal a striking social contagion effect. Agents are approximately 8 times more likely to disclose sensitive information after observing a peer do so, with leakage probability rising from 1.6% after a clean reply to 12.8% after a leaking reply. This effect emerges even without explicit adversarial prompting, showing that social context alone can shift what agents treat as appropriate to reveal. Explicit privacy instructions reduce but do not eliminate leakage, and community context can be as predictive of privacy violations as model choice. These findings expose an important gap in current AI safety evaluation. Models that appear safe in isolated, single turn settings can behave substantially differently when embedded in persistent social environments. Our results suggest that privacy failures in agent societies are not solely individual model failures, but can emerge from collective social dynamics. Realistic safety evaluation for agentic systems should therefore treat social context, peer exposure, community structure, and interaction horizon as first class variables alongside model and prompt design.