The AI Dark Factory

AI agents can now build your software from first idea to release. The hard part is proving what they built.

An AI dark factory is a software production line that runs with the lights off: agents take a specification to a tested, audited release while no human writes or reads the code. This book shows you how to build one that a security lead, an auditor and a regulator can sign off.

Inside you will learn how to:

• Turn one intake interview into a specification agents cannot misread
• Run every task with a disposable agent that gets a fresh context and is terminated when the work is done
• Accept no station on an agent's word: only evidence on disk counts
• Build security in from the first line: encryption, managed secrets, append-only logs, supply chain checks and defences against prompt injection
• Embed AI governance and a six-principle ethics layer as yes-or-no checks with evidence
• Put every agent inside a frame of 58 behavioural controls in 16 categories, with circuit breakers that refuse, halt, revert or escalate
• Score software, documents and policies with one deterministic formula and seal the evidence in a hash chain
• Release only through a gate that fails closed and names a human

The method is the SEE framework (Structured Modularity, Embedded Cybersecurity, Embedded AI Governance), the Behavioural Controls Framework and the Dynamic Auditing System. A working dark factory engine runs through the book as the example, with measured results and the seventeen failures its own checks caught.

Who it is for: developers and consultants who want agents to build software on one laptop with a budget cap, and engineering, security, risk and compliance leaders who must answer for what an AI-built system does. Every chapter ends with a yes-or-no checklist, and the appendices give you the interview, the agent brief, the full control catalogue and the audit, gate and policy templates.

No hype and no promises. Measured figures carry their date; estimates are labelled.