Staff-built AI workflows are spreading faster than IT can track them

Workers are building their own AI-assisted apps and automations faster than security teams can keep up. Some connect to customer records before anyone reviews them. A report in The Hacker News calls this "wild code" and argues that training and policies alone will not rein it in.

What wild code looks like

The example is a sales manager who builds a workflow to route leads. It connects to customer records, summarizes inbound requests and posts updates to a shared channel. It saves time, so the team comes to rely on it. A small experiment becomes a daily dependency before security knows it exists.

This is not just unapproved software. Employees are creating new tools, not only adopting existing products. The risk also goes beyond bugs in generated code. Even code that works as intended can have too much access or show data to the wrong people.

Why the usual fixes fall short

Training, test environments called sandboxes, and AI steering committees all help. But the report says each has limits:

• Training cannot guarantee someone spots the moment they start handling sensitive data.
• Sandboxes protect testing, but a useful prototype eventually reaches real business systems.
• Committees review slowly, may focus on engineering and can miss builds in finance, HR or operations.

Brad Rumph, field CTO at Tines, puts it this way: "Sandboxes solve the design problem and leave the runtime problem alone."

What works better

The alternative is to build controls into the place where employees create and run their work. Security teams set shared boundaries with IT and system owners, apply them throughout a tool's life, and update them as needs change. The report lists several principles:

• Decide which data, actions and access are allowed before systems are connected.
• Give a workflow only the permissions it needs, not everything its creator can reach.
• Require human approval for consequential actions where appropriate.
• Keep building, running and monitoring connected so approved tools do not become blind spots.

A prompt telling an agent what not to do is not an access control. And AI should not decide which business data an app may expose.

What to do

Start by finding out what exists. For each workflow, ask whose credentials it uses, which records it can reach, who can see its output, and whether you could trace and stop a problem. Then make the approved route fast and clear. If it is slow, the report warns, people will use workarounds.