New defense probes detect and mitigate indirect prompt injection in LLMs

Researchers have developed a method to detect indirect prompt injection (IPI) attacks in agentic large language models (LLMs). By training simple linear probes on the models' internal states, they can predict IPI exposure with over 90% accuracy across various models, including GLM-5.2. The study also identified a gap between LLMs' ability to encode IPI signals and their capacity to act safely upon them. To address this, a defense mechanism called AGRI was introduced, which uses probe-gated reasoning to significantly reduce attack success rates while preserving normal task utility. AI

IMPACT This research could lead to more secure agentic LLMs, reducing risks associated with indirect prompt injection attacks.

RANK_REASON Academic paper detailing a new method for detecting and mitigating security vulnerabilities in LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AgentDojo
• arXiv
• GLM-5.2
• indirect prompt injection
• Qwen3.5-27B

• safety
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →