AI in the SOC: Why Autonomy Should Be Earned Action by Action, Not Granted All at Once
Security leaders are preparing to give AI agents far more authority. In an a16z survey of CISOs at organisations with more than 1,000 employees, 36% said they are comfortable letting AI agents act autonomously on low-risk security actions today. That figure is expected to reach 60% within 12 months. The biggest driver is worry that AI-accelerated attacks will outpace response teams, which 55% of respondents selected.
Caution remains at the top end. Only 1% are comfortable with agents acting autonomously on most actions today, and just 7% expect to be in a year. The article argues that treating autonomy as a single slider from "AI suggests" to "AI acts" is why many programs stall. Instead, it points to a 2000 framework by Parasuraman, Sheridan and Wickens, which splits automation into four functions: information acquisition, information analysis, decision and action selection, and action implementation. Each can run at a different level. Teams can automate the first two heavily, where analyst hours go and mistakes are cheap to catch, and stay conservative on the last, where a wrong move disrupts the business.
The piece also highlights a hidden weakness in human approval, which 52% of CISOs say they want for agent actions. If approval requests arrive without clear evidence, reviewers start clicking through to keep the queue moving, and oversight quietly disappears. The suggested test is time: if a reviewer cannot verify a verdict in a minute or two, the approval step is ceremonial.
The article borrows the OWASP idea of least agency, extending least privilege so agents get only the autonomy needed for bounded tasks. Criteria for promotion should be set before a pilot starts, and autonomy should be reduced after a drop in agreement rates, a major environment change such as a new EDR or cloud migration, or a single serious miss. Every extra level of autonomy also raises the damage an agent can do if it is wrong, manipulated or compromised.
Why it matters
Many teams treat human approval as the safety net for AI agents, but this shows the control can fail silently when reviewers lack evidence and simply click through. Security leaders adopting AI SOC tools are pushed to make explicit per-action decisions instead of one trust policy, and to define in advance when autonomy is promoted or withdrawn. It also means each autonomy grant widens the impact if the agent is wrong, manipulated or compromised.
What to do
- Split AI use into the four stages of acquisition, analysis, decision and action, and set a separate autonomy level for each.
- Require every approval request to show enough evidence that a reviewer can verify the verdict in a minute or two.
- Set written criteria for promoting an action class before any pilot begins.
- Demote an agent's autonomy after a drop in agreement rates, a major environment change or a serious miss.