The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents

Summary

Researchers introduce the one-relation attack (ORA), which targets a single implicit warrant linking evidence to a conclusion in LLM agents. Using one source-grounded sentence that qualifies that warrant, best-of-$10 attacks overturned initially correct answers on biomedical inference and vulnerability detection tasks across eighteen models, at mean rates of 54.9% for proprietary and 80.0% for open-weight models. The attacks transferred without victim feedback, and multi-agent debate raised team error from 50.0% to 59.5% in a balanced stress test.

Related items

• HighCVE-2026-77926: Aix-DB remote code execution through ChatBI common agent execute toolSimilar attack · NVD/CVE Database
• CriticalCVE-2026-75128: Langchain-Chatchat unauthenticated shell command executionSimilar attack · NVD/CVE Database
• HighCVE-2026-109443: JoySafeter remote code execution via unsandboxed execute tool in CopilotSimilar attack · NVD/CVE Database
• HighCVE-2026-107366: IBM Langflow OSS message endpoints lack ownership verificationSimilar attack · NVD/CVE Database
• HighCVE-2026-53995: MLflow path traversal in FileStore trace writes via POST /api/3.0/mlflow/tracesSimilar attack · NVD/CVE Database