When Attackers Mimic Your Data: The Critical Flaw in AI Defense Systems

This research exposes a critical vulnerability in multi-agent LLM systems: standard injection detection safeguards collapse when attacks are camouflaged to match domain-specific language, with detection rates plummeting from 93.8%-100% to 9.7%-55.6%. The flaw extends beyond basic detectors to production safety classifiers like Llama Guard 3, revealing that AI systems have a fundamental architectural blind spot when interpreting specialized documents. This threatens the reliability and safety of AI agents deployed in critical domains where they must make consequential decisions based on expert documents.