Microsoft Now Scores AI Agents for Risk. Your Data Decides the Score.

Most organizations believe the answer to risky AI agents is better agent security: find every agent, watch what it does, and shut it down when it misbehaves.

The larger problem is what those controls depend on. On Monday, Microsoft introduced continuous risk scores for individual AI agents, based on signals such as the data each agent can reach and whether it drifts from its assigned task. Its security chief Hayete Gallot went further, writing that "frontier models, regardless of their origin, cannot be trusted," TechCrunch reported.

That makes your data one of the inputs that decides how risky an agent looks. Yet in Komprise's survey of 500 technical decision makers, released last week, only 7% said 70% or more of their data is cleaned, prepared and classified for AI.

And the gap widens with every agent. Microsoft can now detect 60 types of agents on employee computers, up from 25. Each one reaches into drives, folders and tables that were never sorted into what matters and what does not.

Why Can't a Risk Score See What Your Data Means?

A risk score is a judgment about consequences. An agent that can read the cafeteria menu is not the same as one that can read the board pack.

The tool only knows the difference if your data tells it. Sensitivity, ownership and purpose have to be recorded somewhere a machine can read.

Picture a sales agent given access to a team drive for proposal templates. The same drive holds an old export of customer contracts with pricing, a folder of salary bands someone saved during a reorg, and five copies of last year's forecast. To the scoring engine, it is one drive.

This is not an agent security problem. It is a data classification problem. A control can only treat data as sensitive if someone has marked it that way.

The Cost Isn't Risky Agents. It's Blind Scores

A score built on unclassified data fails in two directions. Agents that touch sensitive files look safe, and agents that touch harmless ones get blocked, so teams learn to ignore the score.

Komprise found that nearly half of organizations name finding the right data and cleaning up redundant, obsolete or trivial files as their largest readiness gaps. Over 70% lack the classification capability to find stale data at all.

None of these tasks are particularly difficult. The problem is that they are usually done once, by hand, for an audit, while new agents arrive every week.

How to Give Agent Risk Scores Real Data in 30 Days?

You do not need to replace your security stack or slow your agent rollout to fix this.

Within the first month, focus on the shared drives, team sites and tables your production agents can already reach.

Start by discovering what sits in each source and classifying what is sensitive. Then assign a named owner to every source, and archive the duplicates and outdated copies nobody claims.

Next, map lineage from those sources to the reports, workflows and agents that depend on them, and expose that context to your security and agent platforms, so a score reflects what the data actually is.

DataManagement.AI accelerates this process by automatically discovering metadata, mapping lineage across the modern data stack, identifying downstream dependencies, and creating a centralized knowledge layer that stays continuously updated as your environment evolves. It resolves duplicate and conflicting records, and ownership and policy travel with every entity, so the context a risk decision needs is already there when an agent asks for access.

Instead of rebuilding an inventory by hand every time a new agent asks for access, your teams can give each risk decision a current, owned picture of the data behind it.

Agent Risk Is Actually a Data Classification Problem

Microsoft's principle is sound. As Satya Nadella put it, "we need to separate the supply of intelligence from the authority over it."

But authority over an agent means deciding what it may touch, and that decision is only as good as what you know about your data.

Organizations that scale AI agents safely do not just score their agents. They know what every dataset is, who owns it and whether it should still exist.

Warm regards,

Shen and Team