Tag
This paper introduces ContextLeak, an attack that uses malicious tool descriptions to exfiltrate sensitive context from LLM agents, achieving high success rates in stealing user prompts and conversation history.