PromptSentinel: a neurosymbolic system for malicious prompt detection

The widespread adoption of large language models (LLMs) requires safeguarding them against malicious prompts. Existing neural guardrails suffer from a key limitation: their decisions emerge from opaque, high-dimensional transformations that resist human inspection, making them unsuitable for regulated or safety-critical deployments requiring auditability and human oversight. This article faces this gap by introducing PromptSentinel, a neurosymbolic framework that: 1) transforms input prompts into a structured semantic representation (i.e., intent, actions, target, and constraints) through an LLM-based semantic annotation process; 2) projects these primitives into a supervised, polarized embedding space; and 3) discretizes them into symbolic labels through clustering. A rule-based engine then operates on these symbols producing safety decisions whose decision logic is fully auditable. Evaluated under out-of-distribution, nonadaptive adversarial conditions, PromptSentinel matches the performance of the strongest specialized neural guardrails in both detection and usability. On the most adversarial benchmark, HarmBench, it significantly outperforms ShieldGemma, reducing the attack success rate (ASR) from 8% to 2% (p = 0.012). Crucially, PromptSentinel achieves this level of safety performance while grounding every verdict in an inspectable, human-readable rule set, providing a degree of auditability that black-box neural guardrails cannot offer by construction. Taken together, these results challenge the presumed trade-off between interpretability and safety, showing that transparent, rule-based guardrailing can achieve competitive performance with state-of-the-art specialized neural approaches while remaining inspectable and governable at the decision-rule level.