DREAM: Dynamic Red-teaming across Environments for AI Models

Authors: Liming Lu, Xiang Gu, Junyu Huang, Jiawei Du, Yunhuai Liu, Yongbin Zhou, Shuchao Pang
Published: 2025-12-22

Source: https://arxiv.org/abs/2512.19016

PDF: https://arxiv.org/pdf/2512.19016

AIにより推定されたラベル

脆弱性攻撃手法 動的攻撃評価手法 モデルの堅牢性

※ こちらのラベルはAIによって自動的に追加されました。そのため、正確でないことがあります。
詳細は文献データベースについてをご覧ください。

Abstract

Large Language Models (LLMs) are increasingly used in agentic systems, where their interactions with diverse tools and environments create complex, multi-stage safety challenges. However, existing benchmarks mostly rely on static, single-turn assessments that miss vulnerabilities from adaptive, long-chain attacks. To fill this gap, we introduce DREAM, a framework for systematic evaluation of LLM agents against dynamic, multi-stage attacks. At its core, DREAM uses a Cross-Environment Adversarial Knowledge Graph (CE-AKG) to maintain stateful, cross-domain understanding of vulnerabilities. This graph guides a Contextualized Guided Policy Search (C-GPS) algorithm that dynamically constructs attack chains from a knowledge base of 1,986 atomic actions across 349 distinct digital environments. Our evaluation of 12 leading LLM agents reveals a critical vulnerability: these attack chains succeed in over 70