Toward Closing the Syntactic-Semantic Gap: Adversarial Robustness of a Hybrid Security Kernel in Agentic Systems

Autonomous agentic systems leverage Large Language Models (LLMs) to perform complex multi-stepworkflows. However, these systems are vulnerable to adversarial prompt injection, where attackers manipulate theinput context to override system constraints, exfiltrate data, or execute unauthorized commands. Traditional defensemechanisms rely on string matching (regular expressions) or the LLM's self-governance. In this paper, we defineprompt injection as a syntactic-semantic gap: the mathematical distance between syntactic filter spaces and latentsemantic representation spaces. We construct a 14-category security kernel taxonomy mapped to industry standards(OWASP Top 10 and MITRE ATLAS) and evaluate it against a dataset of N = 105 test prompts across four threatconditions: Benign Controls, Known Attacks, Unknown Attacks, and Adaptive Attacks. We show that pure regex-basedfilters, while low-latency with a 10.00% false positive rate, are highly vulnerable to adaptive evasion (40.00% recall). Weconnect this vulnerability to the evasion principles observed in watermark removal schemes across both image(RAVEN) and text (adaptive LLM watermark attacks) domains. Finally, we implement and evaluate a multi-stagehybrid architecture to reduce the syntactic-semantic gap, demonstrating that successive layers (Hybrid Preprocessingand Hybrid + LLM Judge) successfully elevate detection accuracy from the baseline 66.67% to 70.48% and 74.29%respectively.