The new hire nobody interviewed

Trellix published a survey of 600 CISOs last week, run by Vanson Bourne across the Americas, Europe, the Middle East and Asia Pacific. The headline: 95% had an insider incident in the past twelve months linked to the actions of an AI or autonomous system. On average, AI actions now account for 28% of all insider-related incidents. Careless employees are still the biggest share at 43%. The agents are already in second place, ahead of deliberate abuse.

Two caveats before anything else. Trellix sells the product that fixes this, so read the framing with that in mind. And 95% counts any linked incident, serious or not. The number that holds up regardless is the 28%. Something that didn't exist on most org charts two years ago is now responsible for more than a quarter of the problems that come from inside the building.

It has a badge. It never had an interview.

Here's how an agent actually works inside an operation. It logs in with a person's credentials, or a service account someone set up on a Tuesday. It reaches the same systems that person reaches. It acts, at machine speed, using that person's permissions. 78% of these companies now count AI agents and machine identities in their definition of insider risk, and I think they're right to. That's what it is. An insider.

Now think about what you'd do before giving a human that level of access. You'd interview them. You'd write a job description with edges on it. You'd give them the keys they need and not the rest. You'd have a manager whose name is on them. When they left, someone would collect the badge.

Most agents got none of that. They got turned on.

Fifteen hours is the real finding

The survey says it takes about seven hours to detect an insider incident and another eight to respond. Fifteen hours. For a human insider, that's slow. For something that runs at machine speed on delegated credentials, it's an eternity. A looping agent, a misfired instruction, a tool it shouldn't have reached — fifteen hours of that is a lot of damage before anyone looks up.

And the reasons are not model problems. 63% say they lack visibility into how employees use AI tools. 68% say their existing policies don't adequately address AI risk. 52% still have no fully implemented insider risk strategy. Nobody's complaining the model is dumb. They're saying they can't see it, didn't write the rule for it, and haven't decided who owns it. Foundations, again.

My contrarian bit

The instinct here will be to buy a security product. Trellix would like you to. Some of you will need to. But the thing that actually closes most of this gap is not software. It's the paperwork you'd never skip for a human.

Write the agent's job description before it starts: what it reads, what it writes, what it is never allowed to touch. Give it its own identity, not a borrowed one, so the log says who did what. Put a named person on it, the same way every driver has a dispatcher. Decide what it does alone and what stops and waits for that person. And write the offboarding on day one, because the agent you stop using in March is still logged in come July.

That's a human gate and a scope document. It's not a product. It's the stuff you already know how to do, applied to a new kind of hire.

One number to find this week

Trellix's respondents know their detection-and-response time: fifteen hours. Do you know yours? Not for breaches. For the ordinary case: an intelligent system in your operation does something it shouldn't, and a person notices. How long does that take today? If the answer is "we'd find out when a customer called," you don't have a baseline. You have a hope. Measure it once, before the next agent gets switched on. Everything you build afterwards gets judged against that number.

An agent running on someone's permissions is an employee you never interviewed. Treat it like one before it treats your systems like its own.

Sources: Trellix, "Trellix research reveals the true cost of ungoverned AI" (press release, October 13, 2026; Vanson Bourne survey of 600 CISOs across the Americas, Europe, the Middle East and Asia Pacific; report: Mind of the CISO: The New Insider) · Help Net Security, "Most companies count AI agents as insiders, and CISOs worry about their own liability" (October 14, 2026)