AI alerts in the SOC: up 685%, and 94% of them are noise

Your team adopted AI tools faster than your security tooling learned what they are. Intezer, which runs automated triage across live enterprise alert queues, published numbers this week putting a shape on it: AI-related security alerts rose 685% between February and June 2026. The interesting part is the other number next to it.

What actually happened

Across roughly 16.9 million SOC alerts in Intezer's telemetry, about 73,000 — 0.43% — were AI-related. That is a rounding error on today's volume attached to a slope that steep.

The classification breakdown is the part worth sitting with. Of those AI-related alerts, 94.1% were noise, 5.8% were genuine security risk, and 0.02% were real attacks. Intezer's automated triage returned benign verdicts on 79.8%, suppressed 81.7% outright, and escalated 5.4% to a human. The research was written up by Intezer senior researcher Nicole Fishbein at The Hacker News. These are one vendor's numbers from its own customer fleet, not an industry census — read them as a trend line, not a benchmark.

The alerts are not exotic. They come from ordinary use: developers running coding agents, non-technical staff signing consumer AI tools into corporate accounts, browser extensions, unfamiliar API destinations. Normal work that your detection stack has never seen before and does not have a rule for.

Why this matters for your business

A 94% false-positive rate on a new alert category is the recognizable early shape of a detection gap. Your tooling sees an unknown binary calling an unknown endpoint and fires, because that is genuinely what it should do with no context. The fix is context, not a louder alarm — and nobody is going to supply that context for you.

Here is the part most small companies get backwards. The 0.43% makes this look like a problem for later. It is not, because the noise arrives before the threat does. If you wait until AI alerts are 5% of volume to start classifying them, you will be building the ruleset during the incident. Right now the volume is low enough to sort by hand, which is exactly why now is when it is cheap.

Three moves that cost you an afternoon. Write down which AI tools are sanctioned — the actual list, with the accounts they sign into and the data they can reach. Anything not on that list generating traffic is a finding, and you cannot have findings without a list. Tune before you scale, meaning suppress the known-good agent traffic deliberately instead of letting analysts learn to ignore a whole alert class. Alert fatigue is not a personality trait; it is what happens when a category is 94% noise and nobody prunes it. Watch the OAuth grants, because the highest-risk version of shadow AI is not a chatbot — it is an employee granting a random AI product read access to your Google Workspace or Microsoft 365 tenant in two clicks.

The trend is not going to reverse. Decide now whether your team learns to triage this category or learns to ignore it.

Sources: The Hacker News / Intezer research.