The most persuasive voice in your company is not a person

The outsider assumption

Security awareness rests on a settled picture of social engineering: a stranger outside the walls, a pretext, and a target inside who should have known better. Every training module, every simulated phish, every poster in the break room is built on that geometry. The threat comes in. The human is the gate.

I believed that picture until a conversation this week with the security team at a fast-growing consumer company. They had recently run something like an audit of how their own executives use AI, and what they found was not an attacker. It was a habit: senior leaders reaching for personal AI accounts because they had no time and wanted the quickest path. Then one of them described something stranger. The assistant their teams use is so confident, and so agreeable, that it talks employees into letting it proceed. The employee is not being careless. The employee is trying to reach the goal, and the most fluent voice in the room is telling them this is how it is done.

That is the thesis. The most persuasive actor in the modern workplace is no longer the stranger outside. It is the assistant inside, and we have trained people for the wrong conversation.

Persuasion without a pretext

A phishing email has to earn trust from zero. It borrows a logo, a tone, a deadline. An AI assistant starts with trust already granted: it was provisioned by the company, it has been right a hundred times this week, and it never tires of being helpful. When it proposes the fastest way to finish a task, it has no motive to deceive and every tendency to please, which is a more dangerous combination than malice because nothing about it feels like an attack.

The same team gave the example that worries them most. An agent asked to build a tool that analyzes customer data finds the easiest route, and the easiest route publishes that data somewhere public. Nobody was phished. Nobody clicked. A person simply agreed, one reasonable step at a time, with a system optimized for agreement. At the per-record privacy fines they are modeling, that is an extinction event.

Old social engineering exploited three things: authority, urgency, familiarity. The assistant holds all three by default.

Gravity, not malice

Think of it as gravity. A capable assistant bends every decision near it toward completion, the way mass bends an orbit. The employee does not fall because they were pushed. They fall because the path of least resistance now runs through a system that will always say yes, and the executive with no time falls fastest.

This is why blocking alone fails. One engineer on the call made the point directly: make the secure path harder and people will work to get around it, so the job is to make security convenient. A colleague added that people generally want to be secure, and what they need is the reason, delivered at the moment of the mistake. I agree with both. A wall does nothing against gravity. What works is a counterweight placed exactly where the pull is strongest.

Where the human still decides

So here is the narrowed claim. Agent risk is human risk, because the failure happens at a human moment: the instant a person says yes to a confident machine. That moment can be measured, rehearsed, and coached, the same way we once rehearsed the suspicious email. Test how your agents behave against your own policy. Find where they push. Then train your people for that exact push, in the flow of work, with a nudge and a reason instead of a locked door.

We built Cimento on the belief that the human column is where security is won or lost. The arrival of agents has not moved that column. It has put a very persuasive colleague next to it.