1
2
When Package Managers Can't Help: Defending AI Agent Skills Against Supply Chain Attacks
ClawHub恶意技能占比11.9%，npm防御成熟但AI Agent仍需特殊防护，探讨供应链攻击新战场。
A real-world implementation of static + LLM-based scanning for Claude Code / Cursor skill layers npm's supply chain defenses have matured fast. By 202…
3
Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense
LLM智能体安全新方案：通过自演进对比记忆机制，让防御系统在动态攻击中持续强化，兼具可塑性与鲁棒性。
arXiv:2606.05743v1 Announce Type: cross Abstract: Despite advances in safety alignment, large language models remain vulnerable to continuously evolvi…
4
Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs
提出后门遗忘泛化新路径，让大模型摆脱未知触发器威胁，捍卫LLM安全防线。
arXiv:2606.03785v1 Announce Type: new Abstract: Backdoor attacks in Large Language Models (LLMs) are a growing security concern, where models can gene…
5
SafeMCP: Proactive Power Regulation for LLM Agent Defense via Environment-Grounded Look-Ahead Reasoning
SafeMCP框架通过环境感知前瞻推理主动调节功率，巧妙提升LLM Agent安全性，AI安全研究新突破。
arXiv:2606.01991v1 Announce Type: new Abstract: As Large Language Model (LLM) agents increasingly leverage the Model Context Protocol (MCP) to operate…
6
Aligned but Fragile: Enhancing LLM Safety Robustness via Zeroth-Order Optimization
零阶优化为LLM安全对齐注入抗攻击韧性，揭示脆弱性根源并给出提升方案。
arXiv:2605.29396v1 Announce Type: new Abstract: Safety alignment for large language models (LLMs) aims to reduce harmful or unsafe behavior while pres…
7
Securing Multi-Agent Systems Against Corruptions via Node Contribution Backpropagation
多智能体系统新防御：用节点贡献反向传播精准隔离腐败节点，提升系统鲁棒性
arXiv:2510.19420v2 Announce Type: replace-cross Abstract: Multi-Agent Systems (MAS) have become a prevalent paradigm for Large Language Model (LLM) ap…
8
Trust No Tool: Evaluating and Defending LLM Agents under Untrusted Tool Feedback
LLM Agent面临工具反馈不可信风险，论文提出评估框架与防御方法，为AI安全提供新思路。
arXiv:2605.17453v1 Announce Type: cross Abstract: Tool-using LLM agents increasingly rely on external tools to make consequential decisions, yet most …