AI Dark Code — Organisational Accountability and Control

Narrative

The Substack and practitioner-commentary lane reveals a field rapidly converging on a central thesis: AI-generated and agent-produced code ('dark code') has already escaped the boundaries of conventional enterprise governance, and the management frameworks built for human-authored software are failing in predictable but underaddressed ways. Multiple Substack authors (Secure Trajectories, Rock Cyber Musings, Platforms/AI/BigTech) and academic outlets (California Management Review, Journal of Management Studies, IJRSI) document the same structural breakdown: principal-agent theory, RACI accountability, and periodic human-review oversight all assume intentionality, episodic decision-making, and a human author traceable to every artefact. When production logic has no author — or when the 'author' is an LLM orchestration layer with no intent — these frameworks lose their foundational premise. The most cited empirical flashpoints are Microsoft's admission (April 2025, Satya Nadella) that 20-30% of code in some Microsoft repos is AI-generated with no reliable post-hoc detection method; a Fortune 50 enterprise analysis (September 2025) finding AI-assisted developers produce commits at 3-4x the rate but introduce security findings at 10x the rate; and Veracode's 2025 GenAI Code Security Report showing 45% of AI-generated code contains OWASP Top 10 vulnerabilities regardless of model generation. The governance response is nascent but accelerating: Singapore's IMDA launched the world's first national agentic AI governance framework on 22 January 2026, mandating agent identity management, human accountability checkpoints, and technical controls; NIST launched an AI Agent Standards Initiative in February 2026; and OpenAI, Anthropic, and Block co-founded the Agentic AI Foundation under the Linux Foundation in December 2025, donating AGENTS.md, MCP, and Goose as the first open standards infrastructure for agent provenance and interoperability. On the technical controls side, the emerging consensus centres on: policy-as-code with centralised logging, AI Bills of Materials, pre-commit hooks that threshold AI-generated code at >60% for enhanced review, and OpenTelemetry GenAI semantic conventions for runtime tracing of agent reasoning chains. However, the Rock Cyber Musings Substack post critically identifies that Singapore's human-in-the-loop checkpoint model is arithmetically unscalable for a mid-size enterprise running 50 agents at 20 tool calls per hour — a governance failure mode that the field has not yet resolved.

The observability and discoverability sub-theme shows the most technical maturity: OpenTelemetry's official AI agent framework semantic conventions are now the designated standard for interoperable agent telemetry, with auto-instrumentation packages available for OpenAI, Anthropic, LangChain, and LlamaIndex. Red Hat's April 2026 guide demonstrates production-grade distributed tracing across multi-agent workflows using MCP servers. However, a December 2025 arXiv survey described the agent data landscape as 'increasingly fragmented with loosely defined terminologies' — and Stéphane D.'s Substack documents that agent reasoning data is still being dumped into markdown files and vector databases without structured governance. The governance ownership question remains unresolved: the 2025 Agentic Identity Survey found agent identity ownership split between Security (39%), IT (32%), and an emerging AI Security function (13%), meaning no single function owns the dark-code risk in most enterprises. The unresolved claims in the Substack thesis concern: whether 'policy as code' and signed artefacts can substitute for human authorship as the accountability anchor; whether any enterprise has successfully reversed or paused an AI code generation initiative specifically due to audit failure (as opposed to security incidents); and whether classical stewardship theory — which assumes agent alignment with principal interests — is more applicable than adversarial principal-agent theory for agentic systems that lack intentional self-interest.