"36% said they are comfortable today letting AI agents act autonomously on low‑risk security actions." That finding, from an a16z survey of CISOs at organizations with more than 1,000 employees, sits at the center of a quiet but rapid shift inside security operations centers.
Four automation functions, mapped to the SOC
Longstanding automation research, summarized in a widely cited 2000 paper by Parasuraman, Sheridan and Wickens, breaks automation into four functions: information acquisition, information analysis, decision and action selection, and action implementation. Applied to security operations, those map to four concrete stages: evidence collection (queries to SIEM and EDR, process trees, enrichment and context); analysis and verdict (correlation and reconstruction); response recommendation (what should happen next); and response execution (making changes in the environment). The split clarifies choices: teams can automate evidence collection and analysis heavily where analyst hours are consumed and mistakes are cheaper to catch, while staying conservative on response execution where a wrong move can disrupt production.
Where CISOs stand today — and where they're headed
The a16z survey shows a market moving quickly. Today 36% of surveyed CISOs are comfortable letting AI agents act autonomously on low‑risk actions; in 12 months, 60% expect to be comfortable at that level. AI SOC ranked as the highest‑priority initiative for 2027 in that sampling. But the same respondents remain cautious at the top end: only 1% are comfortable with agents acting autonomously on most actions today, and only 7% expect that a year from now.
Risk is a live concern. The most common worry—selected by 55% of respondents—was that AI‑accelerated attacks will outpace the team's ability to respond. Other top worries included agents acting outside intended permissions or compromised agents (46%) and prompt injection or data leakage through AI applications (29%). The survey also notes current deployment patterns: 55% of respondents run MCP servers or other agent‑to‑application integrations in production, and 34% run autonomous agents with tool access or write permissions.

Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleHuman in the loop vs. human on the loop — and when approvals fail
Enterprise decision frameworks matter. Researchers at the University of Washington describe five levels of agent autonomy tied to the human role, and Gartner distinguishes a human in the loop (approves each action) from a human on the loop (strategic oversight). In the a16z survey, 52% of CISOs want agents to act only with human approval. That sounds safe — until approvals become ceremonial. The article warns human approval is only meaningful when the reviewer can check the AI’s logic quickly. Every approval request should show three things: the evidence the agent collected and its provenance; the reasoning that connects that evidence to the verdict; and what the agent checked and ruled out. If a reviewer cannot verify a verdict in a minute or two, the approval is likely a rubber stamp.
A rubric for decisions: reversibility, blast radius, receipts
Deciding autonomy per action requires an operational rubric. The piece proposes three questions to score each SOC action: How reversible is the action if the agent is wrong? What is the blast radius — how many users, systems or processes are affected? How strong and verifiable are the receipts — the forensic evidence behind the decision? Scoring actions against these criteria produces a map that forces explicit decisions for each action class rather than a single, blunt policy.
Earning autonomy — OWASP guidance and graduation criteria
The OWASP Top 10 for Agentic Applications frames autonomy as something an agent earns. Practical graduation criteria listed include measured agreement (track how often the agent's verdicts match analyst review over a defined period and volume), dry runs that mock writes so teams see intended behavior before changes occur, a complete auditable evidence trail, and a defined rollback path with ownership. Autonomy should also be demoted when agreement drops, when the environment changes materially, or when a serious miss occurs.
How CISOs, AI SOC vendors, and security analysts will respond
- CISOs and security leaders: Inventory every action an AI SOC agent can take, score each against reversibility/blast radius/receipts, automate investigation tasks first, gate disruptive responses with verifiable evidence, and write graduation and demotion criteria before pilots start.
- Enterprises evaluating AI SOC vendors: Ensure vendors support both human‑in‑the‑loop and human‑on‑the‑loop models and provide the forensic‑level evidence necessary to move actions across a defensible trust threshold.
- Security analysts: Expect more automation of evidence collection and analysis — the areas that consume most analyst time — but insist that approval workflows include the evidence, reasoning and ruled‑out checks that make oversight worthwhile.
The article ends with a practical admonition: the projected move from 36% to 60% comfort with autonomous low‑risk actions "will happen with or without a plan." Security leaders who want capacity gains without an incident that sets their program back should follow the five recommended steps: inventory every action; automate investigation aggressively; gate disruptive responses with real evidence; write graduation and demotion criteria now; and govern the agent like a privileged identity (unique identity, scoped permissions, comprehensive logging and detections). Forensic‑level investigation, the article notes, is the approach that moves an action across the trust threshold — a principle Intezer builds into its AI SOC.
Contributed by Zev Schonberg, PMM at Intezer.
https://thehackernews.com/2026/11/the-path-to-ai-autonomy-in-soc.html




