How much control should AI get? A CISO roundtable takes on SOC autonomy

Security teams have spent years drowning in alerts, and AI agents are now being pitched as a way out. The pitch is simple: let machines pull signals from different systems, investigate suspicious activity, and hand humans a cleaner set of next steps. That appeal is obvious when analysts have finite time and attackers are using AI too.

But the shift from help to autonomy is where things get uncomfortable. There’s a big difference between asking an AI to look into a strange login and letting it disable the account behind it. The same goes for isolating an endpoint or blocking traffic. Those actions can happen faster at machine speed, and they can hit the business just as fast if they’re wrong.

That’s why trust is no longer just about the model itself. Security teams need to know what the agent is doing, when a human still has the final say, and whether a bad call can be undone. In practice, that means putting hard limits around autonomy and keeping a way to shut the system down if it goes off course.

The role of the analyst would change too. If AI handles the routine investigation work, people can spend more time threat hunting, making judgment calls, and supervising the agents that are doing the repetitive stuff. The analyst starts to look less like a detective and more like an orchestrator.

The bigger shift may be the SOC itself. AI agents could blur detection, investigation, and response into one continuous loop, with what they learn in one case feeding the next. But there’s a familiar catch: security teams already live with sprawling tool stacks, and vendors are racing to add more agents to them. The promise is continuous security. The risk is just another pile of products to babysit.