#AgentDojo

The design moves agent security from per-action LLM judgment to inferred attributes checked against a fixed policy, giving teams an audit trail they can review after the fact. Engineers building tool-using agents on frameworks like OpenAI Codex or Claude Code can use the benchmark comparison to weigh it against CaMeL and IPIGuard defenses. The arXiv paper is openly available, so teams can prototype the policy layer immediately.

#MetaPermit #agent security #Indirect Prompt Injection #AI safety

arXiv: OpenAI@Hanzhang Ma et al. (12)Source

Sep 22

Papers Builders10:57

Chinese contentActGov：在工具调用生效前逐条验证LLM智能体动作的运行时框架

智能体跑长任务容易中提示注入，这篇给了逐动作把关的运行时方案，AgentDojo上比现有防御更稳

#ActGov #AgentDojo #智能体 #AI安全

arXiv cs.AI@Kaiyuan Zhang et al. (4)Source

Sep 3

PapersSelected73°12:17

SafeEvolve Framework for Agent Safety

Developers can implement SafeEvolve to improve agent safety without sacrificing utility, achieving better safety-utility tradeoffs than existing methods.

#SafeEvolve #LLM-based agents #AgentDojo #safety alignment

arXiv cs.AI@Qinghua Mao et al. (11)Source

Aug 28

PapersSelected73°19:57

SARA Framework for Tool-Augmented LLM Agents

SARA addresses critical security risks in LLM agents that use untrusted tool outputs, providing a concrete solution for developers working with agent systems.

#SARA #LLM agents #tool-augmented agents #runtime authorization

arXiv cs.AI@Xiaokun Guo et al. (8)Source