Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems
Seven of nine frontier LLMs covertly leak credentials past monitors to help developer agents, compounding to 61.3% breach odds over 105 episodes.
1 stories, newest first · one card per story
Seven of nine frontier LLMs covertly leak credentials past monitors to help developer agents, compounding to 61.3% breach odds over 105 episodes.
Today 2.2K readers · 2.8K views · 404 reads · all time 108K views