AI Dark Code — Organisational Accountability and Control

Narrative

The academic and arXiv literature from 2025–2026 reveals a field grappling with a fundamental governance discontinuity: as agentic systems begin autonomously authoring, deploying, and modifying production code, the theoretical and operational frameworks enterprises relied upon—RACI matrices, principal-agent economics, ITIL change advisory boards, model cards—all exhibit structural failures that researchers are only beginning to catalogue. The most theoretically significant work comes from two converging streams. First, management-theory papers (Humberd et al. in the Journal of Management Studies; Jarrahi et al. in California Management Review) apply agency theory to agentic AI and conclude that at the agentic stage, information asymmetry, divergent risk preferences, and goal conflict can exceed what any human-principal monitoring-and-incentive structure was designed to handle. Second, liability papers on arXiv (Xian et al., arXiv 2504.03255) demonstrate that classic legal frameworks for negligent selection and supervision map onto LLM agent deployment only imperfectly, with opacity pushing toward strict product-liability models. On the technical side, provenance and observability emerge as the critical unsolved problem: PROV-AGENT (Souza et al., IEEE e-Science 2025) and the LLM Agents for Interactive Workflow Provenance paper (arXiv 2509.13978) are the first concrete systems extending W3C PROV standards to capture prompt, response, and decision metadata in agentic pipelines—but they remain research prototypes, not enterprise-deployed standards. Governance frameworks have proliferated (the Unified Control Framework, ARGO, TRiSM-for-agents) but empirical studies consistently find an implementation gap: organisations know what governance principles are required but cannot operationalise them at agentic speed and scale. The security data is stark: Veracode's 2025 GenAI Code Security Report found 45% of AI-generated code contains vulnerabilities; agentic AI CVEs grew 255% year-over-year in 2025 per Trend Micro; and the EchoLeak exploit (CVE-2025-32711) against Microsoft Copilot in mid-2025 provided the first documented production incident directly attributable to agentic code execution without adequate governance controls. No academic consensus on a single replacement management framework has emerged; instead, the literature is converging on safety-case and model risk management analogies from aviation and financial services as the most resilient prior art.