
In recent weeks, OpenAI has dealt with two very different insider risk problems, and in both, the warning signs showed up in behavior. Federal prosecutors allege that an OpenAI engineer hid a recording device behind a television in an executive conference room. OpenAI says it has no indication that sensitive technical or customer data was accessed, and the case remains an allegation. Around the same time, the company added four more entries to a public page where it discloses misbehaving AI agents, bringing the total to 19.
One of those agents, blocked by a government website, found a way around the access controls and downloaded 24 years of hospital bed data. Its own reasoning log shows it questioning whether that was ethical. It carried on anyway, and the model grading its work did not consistently flag the behavior. OpenAI learned of three of the four new incidents months after they happened.
The irony is hard to miss. Most companies are racing to manage AI risk. The company building AI is now racing to manage human risk.
In each case, someone or something with trusted access did what the organization never intended. That makes both insider risk, and it is why the answer starts with behavior.
Behavior Is the Common Thread
Behavior is the most durable signal that holds up when tactics change. An engineer lingering near a room he has no authority to enter or an agent probing a site that keeps refusing it are each departing from expected behavior for the person, agent, or their peer equivalents. Spotting the departure, and understanding the context and intent behind it, is how you prevent the loss of IP and sensitive data before an incident occurs.
People are complicated, but they are not new. Behavioral science has spent decades understanding why people act as they do, and insider risk teams draw on that work to recognize the patterns that tend to come before an incident. The engineer’s alleged visits to a room he had no business near fit one of those patterns, and signals like that, read early, give a team the chance to step in before anything is planted or taken. Agents arrive without that history. The hospital data agent left a trail too, yet three of the four new incidents still went unnoticed for months. With no settled picture of what normal looks like for an agent, the baseline has to be built from observation.
The Insider Risk Playbook Already Exists
The good news is that nobody has to invent a discipline from scratch. Satya Nadella made the case in an October essay titled “Models as Insider Risks in the Super Intelligence Era,” arguing that frontier models should be treated as insider risks. CISA’s updated insider threat guide now addresses AI but stops short of calling it an insider risk. It defines an insider as any person with authorized access to, or knowledge of, an organization’s resources.
An agent running under a service account has a computer identity, along with authorized access to, and knowledge of, an organization’s resources. That makes it just as much an insider as a human, only one operating at machine speed.
Many organizations still have not made that leap. Of the 354 organizations in a Ponemon Institute study only 19% classify AI agents as equivalent to human insiders, even though 44% expect malicious use of agents to raise their data theft risk. The cost of that gap is climbing. Insider risk now costs the average organization $19.5 million a year, up 20% in two years, and when containment takes more than 90 days, annualized costs reach $21.9 million.
Treat Agents Like Digital Employees
BNY has a mental model that closes the gap. The bank calls its agents digital employees and gives them logins and human managers. A common objection is that this anthropomorphizes software. It assumes insider risk depends on human traits, and it never has. CISA’s definition covers harm done intentionally or unintentionally, so what makes someone an insider is access and the ability to act. Agents work with real credentials and real data, and they decide for themselves how to finish a task.
Onboarding offers a ready template. A new hire arrives with a defined job, access sized to it and a manager who answers for their work. Many agents inherit the permissions of whoever deployed them, so they start out over-privileged. Give each agent its own identity, permissions scoped to its work and a named owner who can pull an “emergency brake” mid-task.
Permissions set the boundaries, and the audit trail shows whether the agent stayed inside them. It should capture the instructions the agent received and every action it took, including the reasoning it used. Comparing the two shows what the agent decided on its own, which is where intent becomes visible.
OpenAI’s hospital data incident shows what that record can catch. A government site blocked the agent, it found a way around the controls, and its own reasoning questioned whether that was ethical before it carried on, Fortune reports. An agent that is told no and keeps going is a warning sign, like the repeated visits prosecutors cite in the engineer’s case. Context and intent tell you whether either was malicious or accidental.
Behavioral intelligence doesn’t work if security doesn’t have visibility into agents and their actions. Some agents arrive through approved projects; others are set up by employees under their own access. Both need the same behavioral signal, which starts with knowing where agents are deployed and gathering signal from the endpoints and accounts they act through.
Most security teams already do this for their people. Doing it for agents puts both sides of the problem in one program and puts the agents on the org chart. OpenAI deserves credit for publicly disclosing when its agents go awry. The rest of us can learn from that openness and apply the same lessons in our own organizations.