ForgeNEX

AI Agents in the SOC: When to Give Them Real Autonomy

CISOs expect to triple trust in AI agents for security actions within 12 months. Key points for an SMB to delegate without losing control.

4 min read

In this article

The conversation about artificial intelligence in cybersecurity has stopped revolving around whether it should be used and has moved on to a much more uncomfortable question: how much authority are we willing to hand over to it. According to a recent a16z survey of CISOs at organizations with more than 1,000 employees, 36% are already comfortable letting AI agents act autonomously on low-risk security actions. Within twelve months, that percentage rises to 60%. And the SOC with AI appears as the highest-rated initiative looking ahead to 2027.

For a Spanish SMB, the quick reading would be to think this is a matter for large corporations. That would be a mistake. The vendors that today sell MDR and detection tools to large accounts will package those same capabilities for the mid-market within months. The question is not whether they will reach your company, but under what rules you are going to receive them.

What "autonomy in low-risk actions" means

The 36% figure does not mean that a third of CISOs are going to hand over the keys to the SOC to a language model. It speaks to a specific spectrum: blocking indicators, isolating an endpoint, adjusting a detection rule, closing a ticket with sufficient evidence. Reversible actions, with limited impact and, above all, auditable.

The jump to 60% in twelve months indicates that the sector trusts that this autonomy can be bounded with controls. It is a reasonable bet, but it is also a bet: trust in an agent is earned through a track record of correct decisions, not with a sales demo.

Illustrative detail: The Path to AI Autonomy in the SOC

Why this lands in the SMB sooner than it seems

An IT team of five people cannot cover a 24x7 SOC. Nor can it afford to ignore the alerts that arrive in the early hours. The temptation to delegate to an agent that triages and acts on its own is enormous, and it is exactly the terrain where a poorly designed deployment generates more noise than value.

There are three specific pressures pushing in that direction:

  • Lack of hands. Turnover in security roles and the shortage of specialists make every hour of the team worth more.
  • Alert volume. Increasingly cheap detection tools generate more signals than anyone has time to review.
  • Pressure from vendors and insurers. Cybersecurity questionnaires and policies are starting to ask about automation and response, not just about deployed tools.

A practical framework for delegating without scares

Before activating any agent with the ability to act, we recommend setting out in writing these four layers. They are not theory: they are what mark the difference between a controlled pilot and an incident that no one knows how to explain.

  1. Closed catalog of permitted actions. No "the agent decides." The actions it can execute without human approval and those that require sign-off are listed. Everything else is out by default.
  2. Immutable record of decisions. Every action by the agent must be recorded with context: source alert, data consulted, action taken, result. Without a trace, there is no possible autonomy.
  3. Reversibility and damage limits. We prefer actions that can be undone in minutes. Isolating a machine is acceptable; deleting accounts or modifying firewall policies, no.
  4. Periodic human review. A weekly sampling of the agent's decisions detects drift before it becomes habit. Autonomy is renewed, not inherited.

The most common mistake: confusing automation with autonomy

Automating a playbook is writing a rule. Giving autonomy is allowing a system to choose which rule to apply according to the context. The first is audited by reading the playbook; the second requires instrumentation. If your vendor does not show you how the agent's decisions are recorded, you are not buying autonomy: you are buying a black box with access to your network.

It is also worth reviewing the supply chain with the same scrutiny we apply to other components. An agent that queries external APIs expands the attack surface just like any integration. We have already seen how an Anthropic extension in Cursor left a backdoor active for 12 hours; the pattern repeats every time we add pieces that act on behalf of our team.

What we recommend doing this very week

There is no need to wait until 2027 or to hire a managed SOC with AI. With what you already have, you can start:

  • Take inventory of the security actions that a person executes today and that are reversible in less than an hour. That is your natural pilot.
  • Define who responds when the agent makes a mistake. If the answer is "no one," do not activate anything.
  • Ask your vendor for the decision log in an exportable format before signing. If it does not exist, that is a signal.
  • Review the access the agent would have: its own credentials, least privilege and rotation. An agent with the domain administrator account is an incident waiting for a date.

Autonomy in the SOC is going to arrive, and probably sooner than the survey says. The difference between an SMB that takes advantage of it and another that suffers from it will lie in having written the rules of the game before the agent starts playing. As in any fundamental change in IT, the order of the factors matters: control first, speed later.

Source: The Hacker News. Analysis and adaptation: ForgeNEX.

Keep reading

Call us WhatsApp