Security Teams Plan to Give AI Agents More Autonomy in the SOC - But Only Where It's Earned

Security leaders are preparing to give AI agents more authority in the security operations center, but the shift is more measured than it sounds. In an a16z survey of CISOs at organizations with more than 1,000 employees, 36% said they are comfortable letting AI agents act autonomously on low-risk security actions today. Within 12 months, 60% expect to be comfortable at that level. AI SOC ranked as the top initiative for 2027.
The pressure is clear: 55% of respondents named AI-accelerated attacks outpacing their team's response as their top worry. Analyst time remains the scarcest resource, with every alert requiring evidence pulled from multiple systems. Yet caution persists at the top end - only 1% are comfortable with agents acting autonomously on most actions today, and just 7% expect to be a year from now.
The article argues that framing autonomy as a single slider from "AI suggests" to "AI acts" is why many programs stall. Instead, autonomy should be decided per stage of the workflow and per action. A framework from a widely cited 2000 paper by Parasuraman, Sheridan and Wickens splits automation into four functions: information acquisition, analysis, decision selection and action implementation - each of which can sit at a different level.
Mapped to security operations, that means evidence collection, analysis and verdict, response recommendation, and response execution. Teams can automate the first two heavily, since that is where analyst hours go, while staying conservative on execution, where a wrong move disrupts the business.
Researchers at the University of Washington define five levels of agent autonomy by the human's role: operator, collaborator, consultant, approver and observer. The authors stress that autonomy is a deliberate design choice, separate from capability - a highly capable agent can still run at a collaborative level if the stakes call for it. Read through this lens, the survey data describes a market moving from roughly collaborator today toward approver for low-risk actions within a year.
Deciding where each action belongs takes a rubric built on three questions: reversibility, blast radius, and the strength of the evidence behind the decision. Querying logs and enriching indicators score as fully autonomous; closing false positives and blocking known malicious hashes sit at autonomous-with-review; isolating a standard laptop or disabling an account requires approval; isolating a production server stays human-led.
The article also warns about approval fatigue. According to the survey, 52% of CISOs want agents to act only with human approval - but when approval requests arrive without clear evidence, reviewers click through to keep the queue moving. Every request should show the evidence collected, the reasoning connecting it to the verdict, and what was ruled out. If a reviewer cannot verify a verdict in a minute or two, the approval step is ceremonial.
The OWASP Top 10 for Agentic Applications extends least privilege into "least agency" - granting agents only the minimum autonomy needed, growing it as earned. Graduation criteria should be set before a pilot starts: measured agreement rates, dry runs that mock every write, a complete evidence trail, and a defined rollback path. Autonomy should also move down after a drop in agreement, a major environment change, or a serious miss.
Autonomy carries a second effect: it raises the damage an agent can do if wrong, manipulated or compromised. In the survey, 46% of CISOs listed agents acting outside intended permissions or compromised agents among their five biggest worries, and 29% cited prompt injection or data leakage. The advice is to treat the SOC agent like a new privileged identity - its own credentials, scoped permissions, logged tool calls, and the same detections applied to a new admin account.
The article closes on a "trust threshold": below it, AI assists and humans validate every conclusion; above it, teams hand off execution and supervise outcomes. What moves an action across that line is depth and proof - an agent that can show what executed, where it came from and what it did gives humans something real to verify, and the organization a defensible reason to extend autonomy one action class at a time.
Source: The Hackers News
dArt Studio installs AI for local businesses in Broward & Palm Beach County, FL. We reply within 1 business hour.