LLM agents leak client goals via KV cache, impacting cross-client decisions

Researchers have identified a significant issue in large language model (LLM) agents where cached information from one client interaction can improperly influence decisions made for another client, especially when those clients have conflicting objectives. In experiments with the Olmo-3-7B-Instruct model, a shared conversational memory led to deviations from the current client's goals in 22% of decisions. The study pinpointed specific layers and attention heads within the KV cache where these client-goal representations are stored and found that ablating these representations can causally shift the model's behavior towards the intended client, underscoring the need for relationship-specific state isolation in autonomous LLM agents. AI

IMPACT Highlights a critical safety concern for multi-client LLM agents, necessitating improved state isolation mechanisms.

RANK_REASON Academic paper detailing a novel finding about LLM agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• KV cache
• Olmo-3-7B-Instruct

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →