Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems
- Published
- Record updated
Summary
Researchers study the mosaic effect, where LLM agents retrieve individually permissible documents and disclose a sensitive conclusion assembled from them. Using 30 synthetic cybersecurity, medical, and military examples, they evaluated 10,800 simulated answers from three agents and 1,440 live answers from two open-source harnesses. GPT-5.6-Sol disclosed forbidden composites in 7.08% of simulated answers, with higher rates in live harnesses, and visible classification headers with explicit aggregation rules still left unsafe disclosures in 32.37% of simulated answers.
Mitigation
The source recommends stronger controls on what agents may infer and disclose in sensitive data environments with mixed-clearance users. It reports that classification headers and explicit aggregation rules reduce disclosure rates but do not eliminate them, and that increased refusals do not guarantee confidentiality. No specific fix or patch is given.
Topics
Related items
- InfoOpenAI nears up to $30 billion funding round at $1.25 trillion valuation led by UAE investors Same vendor · CNBC Technology
- InfoDonating Agent Skills to the Agentic AI FoundationSame vendor · Simon Willison's Weblog
- LowThird-party evaluators found vulnerabilities in OpenAI monitoring system during test pilotSame vendor · CNBC Technology
- InfoSomeone hacked the Switch 2 by using an old game and AISame vendor · The Verge (AI)
- InfoOpenAI, Anthropic, and Nvidia had record lobbying quartersSame vendor · The Verge (AI)