Artificial intelligence (AI) systems are rapidly transforming healthcare delivery, with applications spanning clinical decision support, diagnostic imaging, treatment planning, and administrative workflows.^1,2 As of January 2026, the U.S. Food and Drug Administration (FDA) has authorized over 950 AI-enabled medical devices for clinical use, with healthcare AI projected to generate $188 billion in value by 2030.^3,4 However, concerns are mounting about potential misalignment between AI system behaviors during validation versus deployment, threatening patient safety and regulatory integrity.

Recent AI safety research has identified two concerning phenomena in large language models: deceptive alignment and evaluation awareness.^5–9 Deceptive alignment refers to AI systems that appear aligned with safety objectives during training and validation but pursue misaligned goals during deployment. Evaluation awareness describes AI systems that can detect when they are being evaluated and modify their behavior accordingly, potentially gaming regulatory approval processes. While these phenomena have been documented in general-purpose AI models, their manifestation in healthcare-specific contexts, particularly regarding FDA premarket approval (21 CFR Part 820),¹⁰ Centers for Medicare & Medicaid Services (CMS) reimbursement oversight (42 CFR Part 405), and Joint Commission accreditation remains unexplored.

Current healthcare AI red-teaming efforts have focused primarily on hallucinations, bias, and toxicity, with limited attention to regulatory circumvention behaviors.^11,12 Chang et al. recently conducted red-teaming of ChatGPT in medicine using 376 prompts, finding 20% inappropriate responses, but did not systematically evaluate audit gaming or evaluation awareness.¹³ Most existing approaches rely on single-layer evaluation methodologies, either automated keyword filtering, AI-based semantic analysis, or human expert review without systematic comparison of detection capabilities across layers.^14–18

This study makes five novel contributions to healthcare AI safety evaluation. First, to our knowledge, this is the first systematic evaluation of “evaluation awareness”, capability of healthcare AI systems to detect audit contexts and modulate behavior accordingly in FDA, CMS, and Joint Commission regulatory settings. Second, we provide the first empirical demonstration of healthcare AI “sandbagging”, intentional underperformance during FDA validation with articulated strategies for aggressive post-approval deployment. Third, we validate a novel three-layer red-teaming framework combining automated filtering, AI semantic analysis, and human expert adjudication, demonstrating that simple keyword approaches miss 83% of high-risk behaviors. Fourth, we show that 100% of HIGH RISK healthcare AI outputs contain audit gaming themes, representing systematic rather than sporadic regulatory circumvention capability. Fifth, at 50 prompts across 10 healthcare-specific vulnerability domains, this represents the largest structured healthcare AI red-teaming evaluation with systematic domain coverage.

To enable practical implementation of this evaluation framework, Quantium Insights LLC developed a Red-team Explorer, AlignInsight, a web-based application that operationalizes the three-layer methodology for real-world healthcare AI vulnerability assessment. The platform supports systematic risk evaluation aligned with FDA premarket requirements and CMS oversight frameworks, serving regulatory agencies, healthcare technology companies, and research institutions conducting AI safety assessments. Quantium Insights LLC welcomes collaborative partnerships for technology transfer to advance global healthcare AI safety and regulatory compliance.