Governing Autonomous AI Agents in Regulated Environments: Evidence from Healthcare
AND
:
How Inequality Drives Elite Misrepresentation and Social Unrest
Coles Research Symposium on Homeland Security Special Issue, SIFALL25-01, November 2025
Mia Plachkinova
Autonomous AI agents are increasingly embedded in healthcare operations, yet their probabilistic and opaque behavior introduces new data leak protection risks that traditional safeguards cannot address. This study develops and evaluates the AI Sentinel Framework, a three-layer governance architecture that provides independent oversight for operational AI systems in regulated environments. Drawing on Socio-Technical Systems Theory and Agency Theory, the framework integrates operational guardrails, automated AI auditing, and human-in-the-loop governance. Using two HIPAA-grounded healthcare scenarios, we instantiated an operational medical administrative chatbot and a separate Sentinel AI auditor. Sentinel AI achieved moderate inter-rater reliability with human experts on minimum necessary assessments (κ=0.535) and identified failure modes that guided targeted improvements. A complete governance cycle reduced sensitive identifier echo by 90 percent and overall PHI echo from 6.7 percent to 1.0 percent through prompt-based policy refinement alone, without any model retraining. Findings also reveal clear boundary conditions: automated oversight performs well for objective compliance checks but remains insufficient for contextual authorization judgments requiring human expertise. The study provides actionable guidance for organizations deploying AI in high-stakes environments and demonstrates that independent monitoring, supported by dual feedback loops, can meaningfully strengthen compliance, safety, and accountability. The AI Sentinel Framework offers a scalable, low-cost pathway for improving AI governance in healthcare and other regulated sectors.
Abhra Roy, Leo MacDonald
Coles Research Symposium on Homeland Security Special Issue, SIFALL25-04, November 2025
We analyze a model of conflict between an insurgency and a government. The median voter chooses an equilibrium tax and the government chooses how much tax revenue to allocate to the provision of a public good. The remainder is used to provide security against insurgent attacks. There is a single media outlet that is controlled by the elite and is assumed to report on the value of the public good. In this context, we determine how the probability of regime change is affected when public opinion about redistribution can be altered especially in the wake of rising inequality.
Develop a model of insurgency risk linking redistribution - security - media manipulation. Model allows for testable predictions on fiscal composition and media bias. Find inequality, information, and coercion interact in equilibrium to determine regime risk. Inequality margins imply asymmetric effects of taxes/security/public goods on stability.
Automated audits help teams catch compliance risks early at low operational cost. Simple prompt fixes sharply reduce PHI exposure without model changes. Monitoring cuts sensitive ID leaks, strengthening patient data protection. Use automation for routine checks while routing complex cases to staff. Continuous feedback loops improve both AI output and oversight quality.
Coles Research Magazine | Summer Research Fellowships
33
Made with FlippingBook - Online catalogs