一作ACL 2026Poster· Natural Language Processing
VIGIL : Defending LLM Agents Against Tool-Stream Injection via Verify-Before-Commit
Junda Lin · Zhaomeng Zhou · Zhi Zheng · Shuochen Liu · an Jiaotong University) · Yong Chen · Enhong Chen (University of Science and Technology of China &
摘要
LLM agents operating in open environments face escalating risks from indirect prompt injection, particularly within the tool stream where manipulated metadata and runtime feedback hijack execution flow. Existing defenses encounter a critical dilemma as advanced models prioritize injected rules du...