Get Started
Home
Topics
Search
Library
Research questionHow can tool-using agents prevent sensitive conclusions assembled from individually non-revealing tool outputs?A tool result may be harmless in isolation yet reveal a sensitive conclusion when combined with other returns and the agent’s reasoning. Per-output privacy checks therefore may miss disclosures that emerge across an agent’s tool-use sequence.
AI Agents
Alignment & Safety
Evaluation & Benchmarks
LLM Pretraining & Post-training
Latest papersRecent research connected to this question, newest first.Agent Tools Orchestration Leaks More: Dataset, Benchmark, and MitigationThe source evaluates this risk under a controlled two-stage tool-use protocol using a 1,000-instance benchmark and six LLM agents. It measures final-response and reasoning-trace leakage, task completion, and H-score, and reports evidence for prompt-based safeguards and post-training mitigation on separate evaluation data.research paper · Sep 2, 2026
Related questions
How can LLM agents reconcile conflicting user instructions, stored knowledge, and live tool observations before acting?How can composable LLM agents preserve authorization and provenance across component boundaries before external effects?How can we train and evaluate LLM agents for tool use across single- and multi-turn workflows with serial or parallel calls?How can LLM-agent systems prevent safety compromises from propagating across workflow boundaries?