Shadow AI in Your Codebase: What Nobody Declared

Most engineering leaders can name the AI tools their company bought. Very few can name the AI capabilities their code already ships.

What "shadow AI" actually means

Shadow AI is any model call, inference library, or vendor SDK that reached production without a review. It usually arrives three ways:

• a transitive dependency that quietly wraps a hosted model API
• a proof of concept that shipped because it worked
• a vendor SDK added for one feature that also phones home

Why it matters

Research consistently puts undeclared dependency exposure between 7% and 20% of a typical codebase. When those dependencies can send data outbound, the exposure stops being a supply chain footnote and becomes a regulatory question.

Finding it

1. Read the code, not the manifest. Manifests describe intent; call sites describe behaviour.
2. Map capabilities, not packages. "Can reach the network" and "can read customer records" are the facts an auditor asks for.
3. Score by reach. A library with filesystem and outbound network access in a payments repo is not the same risk as the same library in a docs site.

You cannot govern what you cannot see. Start with an inventory, not a policy.

Where to start

Run an exposure estimate across your codebase, then scan a single repository to ground the numbers in reality.