AI agents that do tasks correctly leak more private data, a study finds
Leakage increases by a maximum of 3 times from 1 turn to 6 turns.
Claimed, not confirmed
Researchers tested 7 open-weight models from Qwen3 and Llama 3.1 in 4,410 dialogues between 2 AI agents. They found that models that do tasks correctly also leak more private data. No model did tasks correctly and stopped the leakage. The authors write that this is not an error of the test. They write that it is part of how agent systems work.
Sources
Posted
