{"id":8096,"date":"2026-11-05T09:00:00","date_gmt":"2026-11-05T14:00:00","guid":{"rendered":"https:\/\/aventurecloud.net\/?p=8096"},"modified":"2026-11-06T13:35:13","modified_gmt":"2026-11-06T18:35:13","slug":"let-soc-agents-close-the-noise-not-the-attacks","status":"publish","type":"post","link":"https:\/\/aventurecloud.net\/?p=8096","title":{"rendered":"Let SOC agents close the noise, not the attacks"},"content":{"rendered":"

When enterprises ask whether AI belongs in the security operations center, the real question is not \u201ccan it triage alerts?\u201d It is \u201ccan it earn the right to act without handing a false alarm \u2014 or a missed attack \u2014 to a customer?\u201d<\/p>\n

This post looks at a reference build for an agentic AI SOC on Azure patterns: triage, investigation, ATT&CK mapping, threat intel, behaviour analytics, and human-approved containment for a fictional MSSP and three fictional customers. Everything below is measured on synthetic Microsoft Sentinel\u2013style data and scripted attack stories. The language model is a deterministic mock. Nothing was deployed to a live tenant. Treat the numbers as lab evidence, not a production SLA.<\/p>\n

What the project set out to test<\/h2>\n

The build asked three practical questions:<\/p>\n

    \n
  1. Can an analyst feedback loop improve tier routing on the same holdout incidents without ever auto-closing a real attack?<\/li>\n
  2. Can a consensus gate stop false customer auto-reports that a single rule would send?<\/li>\n
  3. Can containment stay behind human approval even when agents draft the plan?<\/li>\n<\/ol>\n

    What was observed<\/h2>\n

    1. Learning cut noise without relaxing the \u201cnever auto-close an attack\u201d rule.<\/strong>
    \nOn the same 40 holdout incidents (days 14\u201320), the baseline without learning scored 77.5%<\/strong> three-class verdict accuracy. After the feedback loop, that rose to 100.0%<\/strong>. Auto-closures moved from 25.0%<\/strong> to 70.0%<\/strong>, human reviews from 30<\/strong> to 15<\/strong>, and simulated analyst minutes from 650<\/strong> to 380<\/strong>. Across both windows and modes, auto_closed_attacks stayed at 0<\/strong>. Earned autonomy here means closing more benign noise, not trusting the agent with an attack verdict alone.<\/p>\n

    2. One rule is a terrible customer reporter.<\/strong>
    \nA consensus gate asked four independent checks (rule, behaviour, threat intel, peer comparison) and auto-reported only when enough agreed, confidence was high, and no check strongly said benign. On 110<\/strong> synthetic incidents with 9<\/strong> true positives, the gate produced 0<\/strong> false auto-reports. A rule-only baseline would have auto-reported 17<\/strong> times with 9<\/strong> false auto-reports. The trade-off is honest: 8<\/strong> true positives still went to Tier 2 because the checks did not agree. That is the cost of not emailing a client a false alarm.<\/p>\n

    3. Detection and containment stayed deliberately asymmetric.<\/strong>
    \nAll 9 of 9<\/strong> attack stories were detected, with a median mean time to detect of 33.0<\/strong> minutes (alert time includes ingestion delay and rule schedule). Containment plans came from a fixed catalogue, approvals were bound to a plan digest, and the executor ran dry only: 0<\/strong> live containment actions. Agents can investigate and propose; they do not pull the plug alone.<\/p>\n

    4. Tenant walls held in the gate checks.<\/strong>
    \nA canary planted in one tenant appeared 0<\/strong> times elsewhere, and 8 of 8<\/strong> cross-tenant access attempts were refused. For an MSSP-style design, isolation is not a slide \u2014 it is a release check.<\/p>\n

    Why it matters for a business<\/h2>\n

    SOC AI pilots often fail in the inbox, not in the model card. A single noisy auto-report burns trust with a customer. An auto-closed ransomware story burns trust with the board. This reference build shows a workable middle path: let agents shrink the queue where the evidence is boring and repeated, force multi-signal agreement before anything goes to a client, and keep every containment action behind a human with an auditable digest.<\/p>\n

    The caveats matter as much as the wins. The holdout is small (one incident moves a percentage by 2.5 points), look-alike benign noise repeats on a schedule so learning is easier than in real life, analysts in the eval are simulated from ground truth, and MTTR is a configured process model \u2014 not a stopwatch from a live shift.<\/p>\n

    Takeaways<\/h2>\n