Agent and Model Detections in Splunk and Sentinel, Without a New Queue

AI agents read files, call tools, and take actions for users. One in eight AI breaches is now linked to agentic systems, according to our 2026 AI Threat Landscape Report. Your SOC sees events your queue has no rules for: a prompt injection that ends in an action, sensitive data leaving through a model. Left raw, they arrive with no technique ID, no owner, and no session tying them together. HiddenLayer runtime detections flow into Splunk and Microsoft Sentinel as records you can route, correlate, and close in the workflows your analysts already use, with no second queue to manage.

The risk also sits below the agent, in the models behind it and in every other AI and ML project in your organization, including model files that can carry known exploits. Model scan results flow into Splunk the same way.

Built for the SIEM you already run

Splunk and Microsoft built their platforms for exactly this job, so we send to them natively. Splunk takes our events through its HTTP Event Collector, which means your searches, correlation, and dashboards apply. Sentinel takes them through Azure Monitor into a Log Analytics table, which means your KQL, analytics rules, and workbooks apply.

In May 2024, Microsoft Azure AI began scanning its model catalog with HiddenLayer. We have also built AI Discovery for Azure, now generally available, and the Sentinel integration.

Splunk holds runtime detections, scan results, and audit events in one place, and Sentinel holds runtime detections and policy outcomes. The HiddenLayer AI Security Platform adds a dedicated view for AI Discovery, AI Supply Chain Security, AI Attack Simulation, and AI Runtime Security when one area needs a closer look.

What an agent looks like in your queue

An agent produces a chain of events, so every record carries the IDs to rebuild it: the requester, the sensor, and a session and interaction ID you can group on in a search. The ordered session, with blocked and redacted turns marked, is one pivot away in the platform. The payloads carry verdicts and identifiers, not prompt text, so your index does not become a copy of your users’ conversations. Agent Harness Security extends runtime security to AI coding agents through an on-device plugin.

A conviction is HiddenLayer’s determination that a malicious event, or set of events, occurred. Each carries a severity, a detection category, and a MITRE ATLAS tactic and technique. Prompt injection is high, data leakage and denial of service are medium, and guardrail activation is low. Incident response and threat intelligence get a technique ID, not a paragraph to decode.

The default policy is built to show you what is there before it stops anything. It detects prompt injection (high) and URLs and source code (low), and it masks emails, phone numbers, IP addresses, and API keys. Agentic Runtime Security produces a record for each. You do not have to alert on all of it. Route high to your queue and low to a hunt index, or build a policy without the rules you don’t want and assign it to the project. When a detection is wrong, report it as a false positive from the detection, which opens a support ticket with metadata attached.

What a model looks like in your queue

Behind every agent in your queue is a model someone pulled in. Scan results tell you who requested it, where it came from, and whether the scan found a known exploit or only a file of concern. Scans separate detections, which are known exploits, from advisories, which are files of concern but not exploits, so a file worth a look does not read like a confirmed compromise.

What to know

Sentinel takes one data source per configuration and supports one HiddenLayer configuration per Azure tenant. Agentic Runtime Security is rolling out by tenant, so confirm availability when you talk to us.

If you want to see what an AI detection looks like in your own queue, book a demo and tell us which SIEM you run.

‍