Domain-Camouflaged Injection Attacks Evade Detection in Multi-Agent LLM Systems

Researchers have identified a critical security vulnerability in large language model (LLM) injection detectors: when malicious payloads are disguised using domain-specific vocabulary and authority structures (called "domain camouflage"), detection rates plummet dramatically—from 93.8% to 9.7% on Llama 3.1 and from 100% to 55.6% on Gemini 2.0 Flash. The study found that even production safety classifiers like Llama Guard 3 failed to detect any camouflaged payloads, and multi-agent LLM systems amplified these attacks, suggesting the vulnerability is fundamental to current LLM architecture rather than a simple implementation flaw.