CISOs to Let AI Agents Act Alone in Security Centres

Security leaders are about to hand AI agents a lot more authority. In a survey of 100 chief information security officers at companies with more than 1,000 employees, conducted by venture firm Andreessen Horowitz, 36 per cent said they are comfortable today letting AI agents act autonomously on low-risk security actions. Within 12 months, 60 per cent expect to be comfortable at that level. The AI security operations centre, or AI SOC, was the highest-ranking initiative for 2027 across the whole survey.
The pressure behind the shift is easy to see. The most common worry among the surveyed CISOs was AI-accelerated attacks outpacing their team's ability to respond, selected by 55 per cent of respondents. Analyst time is the scarcest resource in a security operations centre, and every alert still means pulling evidence from several systems, reconstructing what happened and deciding whether it matters. AI agents promise to take on that grinding work, and the survey shows leaders are ready to let them, at least for the routine parts of the job.
Enthusiasm with strict limits
The same respondents are cautious at the top end. Only 1 per cent are comfortable with agents acting autonomously on most actions today, and only 7 per cent expect to be a year from now. Most teams want more autonomy, just not all of it. More than half, 52 per cent, want agents to act only with human approval, and the survey shows adoption is still early: only 18 per cent run AI SOC tools in production today. But the money is moving. Some 45 per cent plan to buy or upgrade internal AI agent security within 12 months, and 87 per cent expect their security budgets to grow in the 2027 financial year.
The survey, published as "The New Security Agenda" on 28 October, also shows how far AI agents have already spread inside large companies. More than half of respondents, 55 per cent, already run MCP servers or other agent-to-application integrations in production, and 34 per cent run autonomous agents with tool access or write permissions. That is the context for the autonomy debate: agents are not a future experiment for these organisations, they are already wired into live systems, and the question is how much authority they should be given.
Autonomy is a spectrum, not a switch
The analysis accompanying the survey argues that the way most teams frame autonomy is itself the problem. Treating it as a single slider from "AI suggests" to "AI acts" is why so many programmes stall. Instead, autonomy should be split by stage of the workflow and by action. Gathering evidence and analysing it are where analyst hours go and where mistakes are cheapest to catch, so those stages can be automated heavily. Executing a response, such as isolating a production server or disabling an account, carries a much bigger blast radius and should stay under human control for longer.
Three questions do most of the work in deciding where each action belongs: how easily can it be undone, how many users or systems does it touch, and how strong is the evidence behind the decision? A read-only log query can run fully autonomously. Closing a false positive with a full evidence trail can be autonomous with sampling review. Isolating a production server or a domain controller should stay with a human deciding and the agent executing. The point is to make an explicit decision for each action, instead of a single policy that is either too loose for the dangerous actions or too tight for the safe ones.
Approval can become a rubber stamp
There is also a warning about human approval itself. When an approval step arrives without clear evidence or criteria, reviewers start clicking through requests to keep the queue moving, and oversight quietly disappears. If the AI shapes how every investigation is framed, the human approving the final action is approving a conclusion they never saw being reached. Every approval request should therefore show the evidence the agent collected, the reasoning that connects it to the verdict, and what the agent checked and ruled out. A useful test is time: if a reviewer cannot verify a verdict in a minute or two, the approval step is ceremonial.
The survey also underlines that the SOC agent is itself a risk. Some 46 per cent of CISOs listed AI agents acting outside their intended permissions, or compromised agents, among their five biggest worries, and 29 per cent cited prompt injection or data leakage through AI applications. The advice is to govern the agent like a privileged identity: give it its own identity and never shared credentials, scope its permissions to the actions approved at its current level, log every tool call, query and action in a form a human can review, and watch it with the same detections applied to a new admin account. Autonomy, in other words, is something an agent earns with evidence, one action at a time, and it can be taken away again if the evidence stops holding up.
Sources
The Hacker News: The Path to AI Autonomy in the SOC
Vortacity: a16z New Security Agenda TL;DR: CISO Plans for 2027