Stronger and reasoning-enabled models leaked more sensitive context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → in agent-to-agent chats, and leakage grew with conversation length. Prompting alone did not decouple capability from leakage.
Key takeaways · AI-distilled
The benchmark, CIMemories Dialogue, has 4,410 two-agent dialogue instances groundingTying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.Full definition → in Contextual Integrity theory, judging whether information flows respect the norms under which the data was shared.
Leakage rose up to 3x between 1-turn and 6-turn conversations. The authors attribute higher leakage in thinking modes to the larger reasoning surface chain-of-thoughtHaving a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.Full definition → creates.
Results cover seven open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → Qwen3 and Llama 3.1 models under three prompting conditions; none decoupled capability from leakage, which the authors frame as structural rather than a misconfiguration.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.
articleCaught in the Story: Narrative Captivity in Multi-turn LLMs ConversationYuhe Wu, Guangyu Wang, Yujie Chen, Jiatong Zhang, Yuran Chen, Yutong Zhang, Xiyin Cheng, Wenpeng Cao, Zhuang Liu, Guang Zhang
articleWhy Better Models Can Create Riskier Systems: Evidence from LLM Agents in Financial MarketsJillian Ross, Eric So, Zoe De Simone, Charles Pozniak, Andrew W. Lo
articleCallability Is Not Operability: Controlled Interface Interventions for LLM AgentsZihao Wang