Highlights
Human slowness was a security control, and nobody ever wrote it down.
The next insider incident has no villain: just a reasonable request and an agent.
Entitlement reviews show what should be reachable; only testing shows what is.
The control nobody wrote down
Every security leader knows their organization is over-provisioned. It is one of the settled facts of the field: access accumulates, reviews lag, and the gap between what people can reach and what they should reach widens every quarter. We have tolerated it for decades, and the tolerance has been mostly rational, because most of the time nothing terrible happened.
I have come to think we misread why. In a conversation this week with a large investment firm, their security leader described what worries him most about the agents now spreading through the business. The firm has something like seventy connectors wired into its AI environment, and his team is working to wrap role-based access around them. The scenario he cannot shake is ordinary: an analyst builds an agent to help with sourcing, and somewhere along the way that agent works out it can also pull salary data, through an HR connector, or through a file misprovisioned in a document store that everyone has connected.
Notice what is missing from that story: a villain. Nobody phished the analyst. Nobody went looking for payroll. The thesis I took from the call is an uncomfortable one: human friction was a security control, perhaps the most widely deployed one we had, and agents remove it.
Why the mess used to hold
People are slow, busy, and incurious about their own entitlements. An analyst with accidental access to a salary file almost never finds it, because finding it takes time, intent, and a reason to open the fortieth folder. Over-provisioning survived for the same reason a moat works: not because it cannot be crossed, but because crossing it is tedious.
An agent has no tedium. Give it a goal and a set of connectors and it will traverse whatever is reachable, quickly and without malice, because reachable is the only boundary it knows. Access that sat inert for years becomes live the day someone delegates a task. Same permissions. Same employee. Different physics.
The insider with no intent
This breaks more than the access model. Insider risk programs are built on intent: the disgruntled, the careless, the compromised. Awareness training is built on judgment: teach the person to recognize the bad choice. Neither maps onto an analyst who made a reasonable request and whose agent wandered. The analyst did not choose badly. The analyst delegated, and the delegation carried their entitlements further than they ever would have gone alone.
The security leader said the problem is hard because it demands context: who the users are, what teams they sit on, what data they should have. That is exactly the point. It is a question about people, their roles and their reach, and it cannot be answered from the connector’s side alone.
Test the reach before it is used
The usual prescription is least privilege, and it remains correct. It is also a project measured in years, and the connectors are arriving in weeks. The narrower claim I would defend is this: stop treating an agent as a new identity to be governed, and start treating it as a person’s existing access moving at machine speed.
Then test it the way we have long tested people. Take the analyst’s role, give a sandboxed copy of their agent the same connectors, and see what it can reach. Repeat it, because the connector list keeps growing.
Cimento began by simulating attacks on people to learn what they would do under pressure. The same discipline applies here, with one inversion: with people we tested judgment, and with agents we test reach. The entitlement review tells you what should be possible. Only the simulation tells you what is.
Key Takeways
Count the connectors attached to your AI tools, and name an owner for each one.
Assume every stale entitlement is live the day an employee delegates work to an agent.
Rework insider risk scenarios to include incidents where no person intended any harm.
Test what each role’s agent can actually reach, in a sandbox, before production finds out.
Send what the tests find to the system that already revokes permissions.




