Large Language Models (LLMs) have rapidly permeated enterprise, consumer, and governmental applications, fundamentally transforming the human–computer interaction paradigm. However, their widespread deployment has exposed critical security vulnerabilities, most notably adversarial prompt injection attacks, in which maliciously crafted inputs are designed to override system-level instructions, exfiltrate sensitive data, or hijack model behaviour. Existing safeguards, such as coarse-grained output filters and rule-based blocklists, are demonstrably insufficient against semantically sophisticated attack vectors. This paper proposes ShieldLLM—a real-time, hybrid AI firewall that combines Bidirectional Encoder Representations from Transformers (BERT)-derived semantic embeddings with an ensemble Random Forest classifier and a complementary rule-based detection layer to classify incoming prompts as either Safe or Injection Attack with a latency budget under 45 ms. Evaluated on a corpus of 10,000 labelled prompts spanning five injection sub-categories, ShieldLLM achieves 96.3% accuracy, 95.8% precision, 95.7% recall, and an AUC-ROC of 0.982, surpassing all evaluated baselines. The framework is architecturally agnostic and can be integrated as middleware within any LLM serving stack. This work advances the nascent field of LLM-specific intrusion detection and provides a reproducible benchmark dataset for the research community