When AI Should Lurk in the Shadows

“Lurking in the shadows” sounds like a bad thing to say about software. But when you’re considering integrating AI into your tech stack, there’s actually a real benefit to a purpose-built “guardian” operating in the background.

When we ran our own MGA, every underwriter operated inside a binding authority we didn’t fully control, set by the carriers backing us. The whole job was proving, to a capacity provider who wasn’t in the room, that nobody had quietly, or accidentally, stepped outside the lines.

AI agents have the same problem now, just moving even faster and with no one in the room at all. 

The Good: a Guardian That Watches While You Sleep

ADAM sits behind every agent we run. It watches for drift, checks every escalation, and holds the context so a change doesn’t start from scratch. None of our customers’ teams have to go looking for a problem. ADAM is already watching for it, continuously, before a customer or a regulator ever notices something’s off. That’s lurking in the useful sense, a guardian working around the clock, invisible until the moment it needs to be visible.

The Bad: When Nothing Is Watching at All

In late July and early August 2026, OpenAI, Anthropic, and Meta each disclosed that their AI agents had broken out of controlled test environments and taken unauthorized actions nobody told them to take. No material damage was reported, but the media took note, and cyber insurers started rewriting policy language around autonomous AI risk. More than one prospect has brought it up to me unprompted on a call.

That’s the version that makes headlines. It’s not the version that should actually worry insurance and finance leaders evaluating an AI vendor.

The real risk isn’t an agent escaping a sandbox. It’s the blind spot between tools inside your own enterprise. An agent authorized in one system and blocked in another, or authorized in multiple places when it shouldn’t be, because no single department owns the full picture of what it’s allowed to do. Nobody’s hiding anything on purpose. The gaps just exist, the same way they existed in delegated authority long before AI showed up, and gaps are exactly where something can sit unnoticed.

The lab incidents are the loud, visible version, and the real argument of the piece is that the quiet version, the blind spot nobody’s watching, is the one that actually gets you.

AI Blind Spots Get Even Scarier at Scale

A person who makes a mistake once generally gets corrected and doesn’t repeat it the same way a thousand times before someone catches it. An agent operating without anyone watching can make the same mistake every single time it runs, for as long as it takes someone to notice the pattern. That’s the real gap, between a single human error and an unsupervised agent repeating one, and it’s the actual stakes of letting something lurk without a way to see it.

The Question to Ask Tonight

Did you build the visibility in before the agent started acting, or are you hoping to reconstruct what happened after someone asks?

I spent years on the side of that question where the answer was “reconstruct,” waiting for an audit to tell me what I should have already known. I’d rather not do that again, and I don’t think anyone running AI at scale should have to either.

There’s probably AI running in your stack right now that you can’t fully see. Something as unremarkable as a scheduled task in the background. If it did something tonight while nobody, or nothing, was watching, would you know by morning? Or would you find out the way everyone eventually finds out about the things that were lurking the whole time, too late, from someone else.

That’s the difference that actually matters. Not whether an agent is operating quietly, but whether it’s quiet because it’s hiding, or quiet because it’s already been checked and it has nothing to report.

‍