Building an AI triage layer for a small SOC
A design for using a language model to cut alert noise for a two-person security team, and the guardrails that stop it becoming a liability.
A design scenario, not a report on a specific deployment. It describes how I would approach the problem and the controls I would insist on.
AI in security tooling is mostly hype until it removes work a human was actually doing. For a small team fielding a few hundred alerts a day, the work worth removing is first-pass triage.
The problem
Every alert looks urgent until you read it. Analysts spend their mornings closing false positives instead of investigating the handful of things that matter. Alert fatigue is a security risk in itself: the real incident hides in the noise.
The design
A triage layer sits between the SIEM and the humans:
- Enrichment. Each alert is augmented with asset context, recent sign-in history and threat-intel reputation before anyone sees it.
- LLM summarisation. A model condenses the enriched alert into a two-line "what happened / why it might matter" summary and proposes a severity.
- Human in the loop. The model never closes anything. It ranks and drafts; an analyst confirms.
The guardrails
This is where most "AI SOC" projects go wrong:
- The model's severity is advisory. It reorders the queue; it doesn't auto-resolve.
- Every model decision is logged with its inputs, so a wrong call is auditable, not mysterious.
- Alert content is untrusted input. It is clearly delimited in the prompt so a crafted log line can't hijack the summary. (The AI assistant on this site applies the same principle: the server only accepts conversation turns from the client and never a client-supplied system instruction.)
Measuring whether it works
Before switching it on, record a baseline: median time from alert to first human look, false-positive closure rate, and time-to-triage for confirmed incidents. The layer is only worth keeping if real incidents reach an analyst faster than before, not merely if the queue looks shorter.
The aim isn't to replace analysts. It's to give them back the time for the part of the job only a human can do.