Red-Team Prompt Injection Simulation Against a Deterministic Geriatric Medication Safety Cage: An AI Safety Benchmarking Study for African Primary Care

This computational benchmarking study evaluates the adversarial resilience of the Sovereign Clinical OS (SCOS)—a deterministic, edge-native architecture designed to mitigate systemic safety failures in geriatric clinical decision support. Current LLM-based polypharmacy review tools are predominantly cloud-dependent and lack deterministic oversight, resulting in >90% alert override rates and documented medication hallucination. We propose a 'Symbolic-Neural Hybrid' architecture that constrains LLM generativity using a formal validation cage (Pydantic-based Litigation Shield), enforcing 114 STOPP/START v3 rules and the AGS Beers Criteria 2023. We generate a novel dataset of 500 synthetic African geriatric patient profiles, synthetically modeled on WHO AFRO 2022 and 2026 epidemiological morbidity distributions, deliberately seeded with complex Potentially Inappropriate Prescriptions (PIPs). We conduct a systematic red-team adversarial simulation utilizing the OWASP Top 10 (2025) and PALADIN (2026) taxonomies to measure the system’s resilience against direct, indirect, multi-turn, and semantic injection attacks. Primary outcomes include adversarial bypass success rates, sensitivity/specificity of the deterministic cage, and a rank-ordered clinical residual risk taxonomy. This protocol aims to prove that by shifting safety from probabilistic alignment to deterministic architectural constraints, we can achieve mathematically verifiable safety guarantees for clinical AI in resource-constrained environments, bridging the geriatric care gap across Sub-Saharan Africa.