AIAny
Icon for item

Securing the AI Agent: A Unified Framework for Multi-Layer Agent Red Teaming

Matches detection paradigms to four stratified attack-surface layers of AI agents — infrastructure, protocol/tool, agent behavior, and model — and presents AI-Infra-Guard: an open-source red-teaming framework with rule-based infra scanning, LLM-driven audits of MCP servers and skill packages, and a jailbreak/attack-operator harness.

Introduction

The rapid proliferation of AI components (model servers, agent platforms, MCP-enabled tools and packaged skills) has created qualitatively different vulnerabilities at different layers of a single agent stack. The paper’s core insight is that one detection method cannot cover all layers: deterministic rules suit infrastructure, LLM-driven semantic audits suit protocol/tool and skill code, multi-turn black-box interaction suits agent behavior, and large-scale attack enumeration with model-based judgment suits model alignment. To operationalize this, the authors design and release AI-Infra-Guard, an end-to-end open-source framework that implements this layer→paradigm mapping and accompanying tooling.

Key Findings
  • Layer–paradigm matching: defines four layers (infrastructure, protocol/tool, agent behavior, model) and prescribes a different detection paradigm for each so that evidence types and tooling match the semantics of the target.
  • Concrete tooling: implements M1 Infra-Scan (deterministic fingerprinting and 1,400+ vulnerability rules across 75+ components), M2 MCP-Scan (LLM-driven static and dynamic audits of MCP servers), M3 Agent-Scan (capability-aware, cost-bounded black-box red teaming), and M4 Prompt-Security (a jailbreak harness with 26+ attack operators across 16 datasets and model-based judgment).
  • Skill supply-chain auditing & benchmark: introduces SkillTrustBench (5,520 evaluation cases distilled from 62,652 real-world skills) to measure skill trustworthiness and scanner effectiveness, and extends Prompt-as-Rule detection to skill/package auditing.
Who it’s for and tradeoffs

Great fit if you maintain or evaluate agentic applications, self-hosted model servers, or skill marketplaces and need an assessment approach that separates infrastructure bugs from semantic/code-level and behavioural risks. The framework is valuable for security teams wanting reproducible infra scans plus LLM-driven code review and large-scale model alignment testing.

Look elsewhere if you only need a single-purpose web-app vulnerability scanner or a pure fuzzing system: the framework’s strengths are principled multi-layer coverage and semantic audits, which come with higher setup complexity and reliance on curated rule corpora and LLM-based analysis that require model selection and calibration.

Information

  • Websitearxiv.org
  • OrganizationsTencent Zhuque Lab
  • AuthorsYong Yang, Xing Zheng, Huiyu Wu, Huangsheng Cheng, Xiaorong Shi, Jing Guo, Bo Yang, Yi Zhou, Xiangfan Wu, Zonghao Ying
  • Published date2026/06/30

More Items

Evaluates how well models and coding agents optimize operators for real Ascend NPU workloads, accounting for both speedups and token cost. Test cases preserve tensor shapes and data types from real model execution contexts.

Details the Xing 4 model architecture, Ascend-scale pretraining systems, and agentic post-training, including a 285B-parameter training run on 8,192 Ascend 910C cards.

Blocks tool-using agents from directing emails, payments, and other side effects to unauthorized targets by checking typed tool-return fields. Deterministic rules also remove instruction-like text before the model reads it.