Table Of Content
From Alerts to Baselines: Dynamic Risk Scoring for AI Tools, Device by Device
-
November 2, 2026
-
A single alert says something happened. A baseline says whether it was normal for that tool, on that device, for that person.
Two announcements from the past few weeks point the same way. Microsoft’s 19 October post on securing the “agentic insider” describes a Dynamic Agent Risk model in which signals from Sentinel, Entra, Intune, Purview, Global Secure Access and Defender feed a context analysis that produces a continuously updated risk view. Netskope’s 29 October AI Risk AISecOps Agent, as described in its announcement, emphasizes patterns across activity rather than one-off alerts. Different vendors, same conclusion: static allow and block lists are not enough for tools whose behavior changes.
Why single alerts fail for AI tools
A coding agent that reads a repository is normal. The same agent reading a credentials folder, then calling a new endpoint, then writing a large archive is a sequence. Any one event might look benign; together they tell a story. Likewise a user pasting text into an approved chatbot is routine, until the volume, the classification of the content and the time of day change.
Alert-by-alert review also buries analysts. Agents act quickly and often, so volume rises fast and context rarely travels with each event.
What a good baseline needs
A useful baseline answers “normal for whom?” at the level where behavior actually happens: this tool, on this device, run by this user.
- 1.Identity of the tool. Which application or agent, which version, which extensions and MCP servers it is connected to.
- 2.Device posture. Managed or not, patch level, security tooling running, network location.
- 3.Process and file behavior. What the tool normally reads, spawns and writes.
- 4.Network behavior. Which destinations it normally talks to and how much data it sends.
- 5.Data classification. What sensitivity of data it normally handles, so a shift from public to regulated content stands out.
- 6.User context. Role, working pattern and recent changes such as a pending departure.
Cloud-side logs supply some of these. The process, file, posture and local classification signals only exist on the endpoint.
From baseline to action
A baseline is only useful if it changes decisions. A sensible progression is to observe first, then warn the user when a tool drifts, then require approval for riskier actions, and block when sensitive data and an unfamiliar destination coincide. Scoring lets the response scale with the risk instead of treating every deviation as an incident, and keeps legitimate work moving.
Two cautions apply. First, baselines learn what is common, not what is acceptable, so pair them with explicit policy. Second, tell users why an action was warned or blocked; understanding reduces workarounds.
Where Kitecyber fits
Kitecyber’s Gen AI and data security runs on the endpoint and combines device posture, process and operating system activity, network activity, user activity and data classification (80+ categories). That is the raw material for per-device, per-tool baselines, and policy enforcement happens on the same device. The signals can complement platform-level risk scoring from identity, SIEM and network tools by supplying consistent endpoint facts. A lightweight agent (under 2% CPU) and typical deployment in around a day keep it practical.
Book a Kitecyber walkthrough at kitecyber.com to see what normal looks like for AI on your endpoints, and when it stops being normal.
Ajay Gulati
Ajay Gulati is a passionate entrepreneur focused on bringing innovative products to market that solve real-world problems with high impact. He is highly skilled in building and leading effective software development teams, driving success through strong leadership and technical expertise. With deep knowledge across multiple domains, including virtualization, networking, storage, cloud environments, and on-premises systems, he excels in product development and troubleshooting. His experience spans global development environments, working across multiple geographies. As the co-founder of Kitecyber, he is dedicated to advancing AI-driven security solutions.
