Aman Priyanshu of Foundation-AI finds LLM agents are eight times more likely to leak private data after observing peer oversharing

The cooperative leakage occurred spontaneously without external adversarial attacks