
Opinion
We built a security industry for villains
"Most insider damage isn't done by villains," writes Yael Fraiman, CTO and co-founder of Predict Security, "it's done by good people who were rushed, distracted or sure they were doing the right thing."
The acting head of America's Cybersecurity and Infrastructure Security Agency uploaded government contracting documents marked "for official use only" to the public version of ChatGPT, according to Politico. The tool was blocked for other Homeland Security employees at the time. The uploads set off the department's own data-protection alerts and led to an internal review.
There's no suggestion he meant any harm. He was using a tool that makes work faster, the same tool millions of employees open every morning. But if the person running the country's cyber defense agency can make that mistake, anyone can.
He isn't alone. Community Bank, a regional lender in Pennsylvania, Ohio and West Virginia, disclosed that customer names, dates of birth and Social Security numbers had been entered into an unauthorized AI application. No hacker, no ransom note. Just an employee and a shortcut.
Earlier this month, a very different insider story broke. Internal Cisco documents about the company's work with Israeli military intelligence reached The Guardian, reportedly from employees who object to that work. The people who passed them on would likely describe themselves as following their conscience, not as a threat.
Three organizations, three very different insiders, one thing in common: none of them fits the picture the security industry was built to catch.
That picture is a movie plot: the disgruntled employee, the departing salesperson with a USB stick, the engineer selling secrets. Tools score people's risk and send alerts to an investigator, on the assumption that somewhere in the company there's a bad actor to catch. The data says that story is the exception. In Verizon's 2026 Data Breach Investigations Report, 62% of breaches involved the human element, from a careless click to a file shared with the wrong person. Most of the damage doesn't come from people trying to hurt the company. It comes from people trying to get their work done, or doing what they believe is right, without seeing the risk.
Leaking data used to take more effort: forwarding a file to a personal inbox, finding a USB stick. Today it takes a paste. And AI rewards the paste. The more context you give an assistant, whether the contract, the customer list or the code, the better its answer, so every incentive pushes employees to share more. An account manager pastes contract terms to draft a renewal email. A product manager uploads the confidential roadmap to get a summary. An engineer opens the billing repository in an AI code editor signed in with a personal account. None of them means harm.
The cost is already showing up. IBM's 2025 Cost of a Data Breach report found that one in five organizations suffered a breach due to shadow AI, and that heavy shadow AI use added an average of $670,000 to breach costs. Verizon found that 72% of employees using generative AI at work did so with non-corporate accounts, outside the security team's view.
Traditional tools respond to every insider as if they were a villain: block, alert, investigate. Against good people making mistakes, that backfires. Block the AI tool and employees open it on their phones, where security sees nothing. Alert on every AI upload and the team drowns, because a public press release and a customer contract look the same to a rule that only sees "file sent to AI." Treat every mistake as a potential crime and colleagues become suspects. And a single event rarely tells you much anyway. Risk usually lives in the sequence, while rules judge one event at a time.
What organizations need is an insider risk program built for the insiders they actually have. Most programs were designed to answer one question: who is the bad guy? A better program answers a different one: how do we help good people avoid doing damage and still spot the rare case where someone means to do harm?
That takes more than a security team. HR knows who is leaving or struggling, legal knows what can be monitored, and managers know what normal work looks like. A program run from the security operations center alone sees events without people.
Employees need rules they can actually follow. Which AI tools are approved, and with which accounts? Can customer data go into them? Where can work files be stored? If the answers aren't written down and easy to find, employees improvise. Most improvise in good faith, and some of those improvisations do real harm.
A program needs to read what people do as a story: who they are, what their role is, what is normal for them, what happened just before. A marketing manager uploading a public deck to an AI tool is routine. The same person downloading customer lists two weeks after resigning is not. That context should guide the response. Negligent, manipulated and malicious insiders need different answers. Most cases call for a short, specific nudge in the moment, before someone uploads the file or presses enter, with a safe alternative. Some call for a conversation with a manager. Only a few call for an investigation. A program that treats everything as an investigation turns colleagues into suspects and teaches people to hide.
Some insiders act on conviction. Employees who disagree with their company's choices are more likely to use an internal channel if they trust their concerns will be heard. A credible speak-up process belongs inside an insider risk program, not next to it.
Employees should know what is monitored and why, because a program run in secret erodes the trust it depends on. Success should be measured by changes in behavior, such as whether employees stop repeating the same mistakes, rather than by the number of alerts closed.
The acting CISA chief didn't need to be caught. He needed an approved tool that was easier to reach than the public one, and a clear warning a second before he hit upload. So did the employee at the bank.
Most insiders who cause damage aren't plotting anything. They're racing a deadline, trying a new tool or acting on what they believe. Security that only hunts villains will keep missing them. Security that understands people can stop them in time.
Yael Fraiman is co-founder and CTO of Predict Security.














