Security leaders are about to hand AI agents a lot more authority. In a recent a16z survey of CISOs at organizations with more than 1,000 employees, 36% said they are comfortable today letting AI agents act autonomously on low-risk security actions. In 12 months from now, 60% expect to be comfortable at that level. Overall, AI SOC was the highest ranking initiative for 2027.

The pressure behind the shift is easy to see. In the same survey, the most common worry was AI-accelerated attacks outpacing the team's ability to respond, selected by 55% of respondents. Analyst time is the scarcest resource in the SOC, and every alert still means pulling evidence from several systems, reconstructing what happened and deciding whether it matters.

The same respondents are cautious at the top end. Only 1% are comfortable with agents acting autonomously on most actions today, and only 7% expect to be a year from now. Most teams want more autonomy, just not all of it.

Discussions today frame autonomy as a single slider that runs from "AI suggests" to "AI acts," and that exact framing is why so many programs stall. Autonomy is a set of decisions made per stage of the workflow and per action, and the teams that get it right treat it as something an agent earns with evidence.

Autonomy is not “all or nothing”, but rather a spectrum

The question "how much do we trust the AI" is too broad to answer. Researchers worked this out decades ago, in a framework later applied widely in aviation and air traffic control, long before anyone was building SOC agents.

In a widely cited 2000 paper, Parasuraman, Sheridan and Wickens proposed that automation applies to four broad functions. Those are information acquisition, information analysis, decision and action selection, and action implementation. Each function can sit at a different level of automation, from fully manual to fully automatic, within the same system.

Mapped onto security operations, the four stages look like this.

  1. Evidence collection. Querying the SIEM and EDR, pulling process trees, enriching indicators, gathering identity and cloud context.
  2. Analysis and verdict. Correlating what was collected, reconstructing what happened and deciding whether the alert is a real threat.
  3. Response recommendation. Proposing what should happen next, such as containing a host, resetting credentials or closing the case.
  4. Response execution. Actually changing something in the environment.

This split gives security leaders a cleaner set of choices. A team can automate the first two stages heavily, since that is where analyst hours go and where mistakes are cheapest to catch. The same team can stay conservative on the fourth stage, where a wrong move disrupts the business. As the a16z report itself puts it, gathering logs is different from changing a production system.

Human IN the loop vs. Human ON the loop

Once autonomy is split by stage, teams need a shared way to describe how much of it each stage gets. A useful vocabulary comes from researchers at the University of Washington, who define five levels of agent autonomy by the role the human plays.

Level Human role What it means in practice
L1 Operator The human directs and decides, and the agent carries out instructions
L2 Collaborator Human and agent plan and execute together
L3 Consultant The agent leads and checks in for expertise or preferences
L4 Approver The agent works independently and asks for approval only in risky or predefined cases
L5 Observer The agent acts on its own while the human monitors and can hit the stop button

The authors make a point every SOC leader should take to heart. Autonomy is a deliberate design choice, separate from how capable the agent is. A highly capable agent can still run at L2 if the stakes call for it.

Read through this lens, the a16z data describes a market moving from roughly L2 today toward L4 for low-risk actions within a year. Gartner offers a complementary distinction between a human in the loop, who approves each action, and a human on the loop, whose oversight is strategic.

Enterprises evaluating AI SOC vendors should make sure that both models can be applied as appropriate, across their entire operations. 

How to decide where and how much autonomy to grant an AI SOC

The levels mentioned above tell you what the options are. Deciding where each SOC action belongs takes a rubric, and three questions do most of the work.

  1. Reversibility. If the agent gets this wrong, how quickly and completely can it be undone?
  2. Blast radius. How many users, systems or business processes does the action touch?
  3. Receipts. How strong and verifiable is the evidence behind the decision? An action backed by forensic proof of what was executed deserves more latitude than one backed by an LLM hunch about what looks suspicious.

Scoring real actions against those questions produces a map that most security teams will recognize.

Example action Reversibility Blast radius Suggested starting level
Query SIEM, EDR or identity logs Nothing to reverse None L5, fully autonomous
Enrich indicators and pull process trees Nothing to reverse None L5, fully autonomous
Close a false positive with a full evidence trail Easy, the case can be reopened Low L4, autonomous with sampling review
Open and route a ticket to an owner Easy Low L4, autonomous with sampling review
Block a known malicious hash or domain Easy Low to medium L4, autonomous within defined policy
Isolate a standard employee laptop Moderate One user L3 or L4, approval for sensitive groups
Disable an account or revoke sessions Moderate One user, possibly many if a service account L3, approval required
Isolate a production server or domain controller Hard Business-wide L2, human decides and agent executes

The exact placement will differ by organization, and that is the point. The rubric forces 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.

When human approval becomes a rubber stamp

According to the a16z survey, 52% of CISOs want agents to act only with human approval. It sounds like the safe choice, and in many cases it is. It also hides a failure mode that rarely gets discussed.

Human approval is only as good as the context the reviewer receives. 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.

Only when the reviewer can actually check the AI’s logic and work along with hard, forensic evidence that conclusively points in one direction or another, is the approval actually worth anything. 

In practice that means every approval request should show three things.

  • The evidence the agent collected and where it came from.
  • The reasoning that connects that evidence to the verdict.
  • What the agent checked and ruled out along the way.

A useful test if this is working, is time spent. If a reviewer cannot verify a verdict in a minute or two, the approval step is ceremonial and the team is paying for a human-in-the-loop without getting it.

Autonomy you earn, autonomy you lose

The OWASP Top 10 for Agentic Applications gives this approach a name. It extends least privilege into least agency, granting agents only the minimum autonomy needed for safe, bounded tasks and letting that autonomy grow as it is earned.

Earning it requires criteria that are set before the pilot starts. Otherwise autonomy expands on vendor promises and gut feel. A practical set of graduation criteria for moving an action class up one level looks like this.

  1. Measured agreement. Track how often the agent's verdicts match analyst review over a defined period and alert volume, and set the threshold in advance.
  2. Dry runs first. Run new actions in a test mode that executes the logic but mocks every write, so the team sees what the agent would have done before it does anything.
  3. A complete evidence trail. Every decision at the new level must be auditable after the fact, including what the agent looked at and why it concluded what it did.
  4. A rollback path. Define how the action is undone and who owns that before expanding scope.

Autonomy should also move down. A drop in agreement rates, a major change in the environment such as a new EDR or cloud migration, or a single serious miss should trigger a demotion until the criteria are met again. Teams that only ever ratchet upward are building risk they cannot see.

Your SOC agent is an agent too

Every level of autonomy granted to a SOC agent has a second effect. It raises the damage that agent can do if it is wrong, manipulated or compromised.

Security leaders already feel this risk across the enterprise. In the a16z survey, 46% of CISOs listed AI agents acting outside their intended permissions, or compromised agents, among their five biggest worries, and 29% cited prompt injection or data leakage through AI applications. Meanwhile 55% already run MCP servers or other agent-to-application integrations in production, and 34% run autonomous agents with tool access or write permissions.

The lesson for the SOC is that autonomy itself is a risk factor, scored alongside the vulnerability. Treat the AI SOC agent the way you would treat a new privileged identity.

  • Give it its own identity, never shared credentials.
  • Scope its permissions to the actions it is approved for at its current level.
  • Log every tool call, query and action in a form a human can review.
  • Watch it with the same detections you would apply to a new admin account.

A playbook for the next 12 months

The move from 36% to 60% will happen with or without a plan. Security leaders who want the capacity gains without the incident that sets their AI program back a year can start with five steps.

  1. Inventory every action. List everything your AI SOC agent can do today, from read-only queries to response actions, and place each one on the reversibility, blast radius and receipts rubric.
  2. Automate investigation aggressively. Evidence gathering and analysis consume most analyst hours and carry the least risk. This is where full autonomy pays off first.
  3. Gate disruptive response with real evidence. Keep approval on high-impact actions, and make sure every approval request carries proof a reviewer can verify quickly.
  4. Write graduation and demotion criteria now. Agree on the accuracy thresholds, test periods and rollback paths before the pilot starts, so autonomy expands on data.
  5. Govern the agent like a privileged identity. Scope it, log it and monitor it, because its autonomy is also its blast radius.

Crossing the trust threshold

Every SOC that adopts AI agents will eventually face the same question for each action. Is the agent's work trustworthy enough to act on without a human checking it first?

A helpful way to think about it is a trust threshold. Below it, AI assists analysts, and humans still validate every conclusion. Above it, teams can safely hand off execution and supervise outcomes instead of steps. What moves an action across that line is depth and proof. An agent that can show what actually executed, where it came from and what it did, gives a human something real to verify, and gives the organization a defensible reason to extend its autonomy. Forensic-level investigation, the approach Intezer’s AI SOC is built on, is what moves an action across that line.

That is what earned autonomy means in practice. The agent gets more authority one action class at a time, each step backed by evidence the team can inspect, and each step reversible if the evidence stops holding up.

Note: This article has been written and contributed by Zev Schonberg, PMM at Intezer.

Found this article interesting? This article is a contributed piece from one of our valued partners. Follow us on Google News, Twitter and LinkedIn to read more exclusive content we post.