Who owns the risk when an AI agent runs the security desk?

The question nobody can answer yet

A security operations center, or SOC, is the team that watches a company's systems for signs of attack. Many now say they are powered by AI. As CISO Mag reported from an episode of EC-Council's The Cybersecurity Podcast, Jeremy Hehl of Foresite Cybersecurity was asked what that looks like in daily use.

His answer starts with a blunt question he puts to customers who want a fully autonomous SOC: if something goes wrong, who gets fired? You cannot really dismiss a piece of software, so a person has to hold the responsibility.

The piece is a conversation, not a news announcement. It is useful because it shows where a working provider draws the line.

Agents compute, humans decide

Hehl says agents can already do much of the investigative legwork: gathering data, adding context, and linking related alerts into single cases. He says they do this work well. What they should not do alone is decide whether a problem matters to the business. That is the risk call.

He compared the agent to a tricorder, the handheld scanner from science fiction that characterizes a threat instantly so the human can choose a response. The host summed it up as agents compute and humans decide, and Hehl agreed.

He added that the few vendors promising fully autonomous operations are narrow and tightly focused. Agents that both reach a decision and carry out an action need rigorous testing and simulation. Without clean data, he said, they hallucinate, meaning they produce confident but wrong output.

Taming software that gives different answers

The host raised a real engineering problem: ask a leading model the same question five times and you may get five answers. Hehl's reply was to roll out in stages. Start with narrow jobs where machines clearly beat people, such as running queries, calling other systems and writing code.

In enterprise projects, Foresite first gets data normalization, which means tidying logs into consistent fields, to what he called basically four nines. Only after that do agents enter the automated response playbooks.

He also argued for tight limits on each agent's identity, so it can do only very prescriptive things instead of everything. Treating agents like privileged users, with narrow permissions, is the thread running through his advice.

Jobs, costs and the missing middle

On jobs, Hehl said Foresite has not eliminated any; it has grown. With agents helping, he said, analysts handle 10 to 15 times the workload, dismiss false alarms faster and spend the saved time on more meaningful work. He compared this to the Industrial Revolution, where more productive workers led owners to demand more output.

Cost is a sticking point. Token-based pricing, where you pay for the small chunks of text a model processes, is not yet predictable. Hehl said Foresite once ran an accidental $250,000 query on Gemini during testing and had to ask Google for forgiveness.

The host described a large middle tier of companies that know AI is more than chatbots but cannot afford top-end oversight tools. Hehl's answer for them is partnership with managed providers that supply governance and observability as part of the service.

Governance is the next hurdle

Hehl called governance the next real frontier for managed security providers. He named ISO/IEC 42001, an international standard for managing AI, as the emerging benchmark in the US. The host noted it has more than 190 controls and that holding a certificate is not the same as living by it. Asked whether firms show compliance while running with gaps, Hehl said it happens all the time.

He expects the next push to be identity and access management for agents, treating them as users. His advice is to move telemetry, the raw activity logs, into a major cloud, where he says policies are native by default and agents are governed by default.

His forecast: removing human control over security posture is unlikely in the next decade or so.