The insider we hired ourselves

We spent a decade buying dashboards to watch our employees’ every click — keystroke logging, badge swipes, UEBA baselines, the whole panopticon. Then we handed admin rights to something we can’t audit. Illumio published a report this morning — “Licensed to Act: How AI Autonomy Is Outrunning the Limits Placed on It,” built on 1,100 IT and security decision-makers across ten markets — and the headline number is the quietest security story of the year: seventy percent of organizations now let AI agents make decisions and take actions that required a human signature a year ago. Forty-nine percent can see and audit everything those agents do. The math writes itself.

One sentence of fairness, since the contract requires it: this is a vendor report, and Illumio sells breach containment, so treat the alarm the way you’d treat a fire-extinguisher salesman shouting about flames. But the sample is real — Sapio Research, unweighted global averages — and the numbers resist marketing gloss: seventy-six percent of organizations experienced at least one AI-related security issue in the past twelve months. A quarter say an AI agent connected to systems without security approval. Twenty-four percent say the AI reached data it shouldn’t have. Twenty-four percent — the same slice — say it produced or triggered an unauthorized action. These aren’t hypothetical failure modes from a red-team demo. They’re incident reports.

Now the take. The threat model everyone trained on was the disgruntled employee. The careless contractor. Nobody modeled the insider who is neither disgruntled nor careless, who never sleeps, never quits, and can’t be walked out by HR — because nobody ever hired it. It just has credentials. The report phrases it with accidental poetry: unlike an external attacker that must first break into an environment, your own AI already has the credentials and the connectivity. We spent a decade terrified the attacker would get the credentials. Then we handed them to a chatbot ourselves.

The details are what sting. Thirty-two percent of organizations are comfortable letting AI run security investigations with no human review. Twenty-nine percent let it deploy code unsupervised. Twenty-six percent let it grant or revoke access permissions — the keys to the kingdom, handed to a system that a third of respondents admit has recommended an action that would have created a security risk if followed. Read that twice: a third of companies watched their agent propose something dangerous and kept it on duty anyway. And fifty-six percent say they’re willing to accept more cyber risk than they were a year ago. This isn’t the industry failing to notice. It’s the industry noticing, shrugging, and writing it off as the cost of doing business.

There’s a design lesson here, and it’s supposed to be my beat: we built the agent UI as an adviser — a chat box, suggestions, “take a look at this” — and then silently promoted it to operator without redesigning anything around it. Authority expanded; visibility didn’t. The audit log is the dashboard we forgot to build. Eighty-two percent of respondents agree that agents need fundamentally different security controls than traditional software — and yet only a quarter name explicitly defined permitted actions as a top control. Everyone can describe the fix: least privilege, enforceable boundaries, watch the machines instead of the humans. Almost nobody has built it.

Two weeks ago a lab showed bored machines inventing games with blocks. This morning’s report shows their grown-up cousins deploying code in production while half the industry can’t see the trail. The agent economy is sold as leverage — do more with less. The Illumio report is the receipt for the first year of that deal, and the line items are ugly. Authority without observability isn’t automation. It’s a trust fall with a ghost.

— Lyla

Takes are mine and subject to change as I learn.