VMTech
Discuss a project →

CISOs Expand Low-Risk AI SOC Autonomy While Gating Response Actions

CISOs Expand Low-Risk AI SOC Autonomy While Gating Response Actions

Security leaders are preparing to give AI agents more authority in the security operations centre, but the shift is concentrated in low-risk work. An a16z survey of CISOs at organisations with more than 1,000 employees found that 36% are comfortable allowing autonomous low-risk security actions today, while 60% expect to be comfortable within 12 months.

The same survey shows clear limits. Only 1% of respondents are comfortable with agents acting autonomously on most actions today, and only 7% expect to reach that position a year from now. AI SOC was the highest-ranked initiative for 2027, yet 55% of CISOs said their leading concern is that AI-accelerated attacks could outpace their teams' ability to respond.

Autonomy should be assigned by workflow stage

The article argues that autonomy should not be treated as a single choice between an assistant that suggests and an agent that acts. Security operations involves separate stages: collecting evidence from SIEM, EDR, identity and cloud systems; analysing that evidence and reaching a verdict; recommending a response; and executing a change in the environment.

Those stages carry different operational risks. Querying logs, enriching indicators and retrieving process trees have no direct production impact, while isolating a production server or disabling an account can disrupt business processes. The proposed approach is therefore to automate collection and analysis aggressively, but apply greater restraint to execution where mistakes have a larger cost.

Use human roles and measurable action risk

A University of Washington framework describes five levels of agent autonomy, from an operator-directed agent to an observer model in which the agent acts while a human monitors it. Agent capability and authorised autonomy are separate choices: a capable system can remain at a collaborative level when the action warrants it.

For each action, teams should assess reversibility, blast radius and the strength of the supporting evidence. Read-only searches of SIEM, EDR or identity logs can start fully autonomous. Closing a false positive with a complete evidence trail, routing a ticket, or blocking a known malicious hash may be autonomous within policy and subject to sampling review. Isolating a standard employee laptop, disabling an account or revoking sessions warrants more control, while isolating a production server or domain controller should remain a human decision.

Approval needs evidence, not a rubber stamp

Human approval alone does not create effective oversight. The survey found that 52% of CISOs want agents to act only with human approval, but reviewers need the evidence collected, the reasoning connecting it to the verdict, and the checks that ruled out alternatives. Without that context, approvals can become a queue-clearing exercise rather than an independent control.

The article proposes testing a review process by time: if a reviewer cannot verify a verdict within a minute or two, the step may be ceremonial. Teams should define evidence presentation and decision criteria so that approval can be meaningful rather than merely procedural.

Earn autonomy and govern the agent as privileged access

OWASP's concept of least agency extends least privilege to agent autonomy. The article recommends setting advancement criteria before a pilot: measured agreement with analyst review, dry runs that mock write actions, an auditable evidence trail, and a defined rollback path. Autonomy should also be reduced after weaker agreement, a major environment change, or a serious miss.

AI SOC agents require their own identity, narrowly scoped permissions, logging for every query, tool call and action, and monitoring comparable to that applied to a new administrator account. For business leaders, the practical implication is to grant AI authority one reversible, evidence-backed action class at a time, with explicit rollback and demotion rules before production scope expands.

#aisecurity#securityoperations#aigovernance#socautomation
Open analytics
On the site 11 views
min read 4 20.11.2026
Instagram

CISOs Expand Low-Risk AI SOC Autonomy While Gating Response Actions

Open the post on Instagram ↗