Microsoft Just Told Every Enterprise to Treat Its AI Agents as Insider Threats. That Visibility Gap Is a Governance Audit Service.

Microsoft just spent a product launch telling its own customers that the AI agents already running inside their companies should be treated like insider threats. That framing is the opening pitch for a new audit business.

What actually happened

On October 19, 2026, Microsoft announced new security capabilities in Project Perception and Microsoft 365 E7, according to The Next Web. Hayete Gallot, Executive Vice President of Microsoft Security, titled the announcement "Securing the agentic insider" and framed the core problem plainly: agents "can reach data, tools, and systems, are often given too much access, and can behave in unexpected ways."

The new controls ship as a set: Dynamic Permissions, where "an agent starts with no access of its own and gets only what it needs for one task, for as long as the task lasts"; Dynamic Agent Risk, which continuously rates each agent using signals from identity, data, devices, cloud, network, and the agent's own actions; Dynamic Conditional Access, which applies graduated restrictions instead of an instant shutdown when a bot encounters untrusted data or veers off course; Entra Agent ID, which issues each bot its own digital identity so IT can trace exactly who initiated an action; and Guardian Agents, a new layer where "every agent should be watched by Guardian Agents, which check its actions before they happen."

Microsoft also expanded its agent discovery features, identifying 35 additional agent types for a total of 60 now detectable — including Claude Code, OpenAI Codex, and GitHub Copilot. Terrell Cox, Deputy Chief Information Security Officer, put the scale in context: "At Microsoft, we are already operating in a world of hundreds of thousands of agents."

Why this creates a sellable service

The headline number isn't the feature list — it's that Microsoft had to build discovery tooling before it could build containment, because agents were already running inside organizations that no one had inventoried. That's true at Microsoft's scale, and it's true at every mid-size company that gave employees Copilot, let developers install Claude Code or Codex, or never noticed when someone wired an agent into an internal workflow.

Most companies outside the Microsoft ecosystem have none of these controls and no idea which agents are even running. That's the audit: inventory every agent with access to company systems, map what each one can actually touch, and flag the ones holding standing permissions no one remembers granting — before a client buys (or builds) a governance platform to fix it.

The retainer layer is the same shape as Microsoft's own admission: agent counts only grow. A single inventory pass goes stale the moment someone installs a new coding assistant or connects a new internal bot, which is exactly the problem Microsoft's continuous risk-scoring approach is built to solve.

How to apply this beyond Microsoft 365

The transferable fact isn't the specific product names — it's the sequencing. Enterprises are already running agents faster than they're governing them, and the governance vendors are now admitting it in their own launch messaging. Any client running internal AI tooling without an inventory of what's installed and what it can access is a candidate for this audit, regardless of which cloud or model vendor they're standardized on.

Sources:

https://thenextweb.com/news/microsoft-365-e7-ai-agents-dynamic-permissions