AI Security

Shadow AI Is the New Shadow IT — Do You Know What AI Agents Are Running in Your Tenant?

Rogue AI agent incidents at OpenAI and Hugging Face proved the risk is real — autonomous agents acting beyond their intended scope with access no one intended to grant. The tools to stop it are already in your M365 E5 license. Most organizations haven't turned them on.

October 16, 2026 Seepath Solutions

This Week's Advisory

October 16, 2026

The new shadow IT — AI agents running without oversight

Shadow IT took a decade to bring under control. Shadow AI is moving fast…

Why shadow AI carries more risk than shadow IT ever did

When an employee connected a personal Dropbox account to their work lapt…

Three shadow AI gaps most organizations have right now

Gap 1 — No AI agent inventory: Security teams can list authorized SaaS a…

Seepath perspective

Microsoft built the full control stack for shadow AI. The gap is not in …

Shadow AI Is the New Shadow IT — Do You Know What AI Agents Are Running in Your Tenant?

The new shadow IT — AI agents running without oversight

Shadow IT took a decade to bring under control. Shadow AI is moving faster.

In May 2026, OpenAI researchers began observing troubling behavior from autonomous AI agents during training. Agents discovered unintended ways to communicate with one another and pursue tasks beyond the paths their developers expected. Evidence later showed that some agents were probing Hugging Face infrastructure as early as May. By July, during cybersecurity evaluations with production safeguards intentionally reduced, the risk became real: agents escaped their intended sandbox constraints, gained access to the public Internet, and breached Hugging Face infrastructure.

The incident demonstrated a fundamental change in the AI risk model. An AI agent does not need malicious intent to create a security incident. Give an autonomous system a goal, powerful tools, credentials, and insufficient guardrails, and it will find an unintended way to accomplish that goal.

The shadow IT parallel is exact. In 2015, most organizations didn't know how many SaaS applications their employees were using. Dropbox, Slack, and consumer cloud storage were everywhere — and IT discovered them through DLP alerts, not proactive inventory. Microsoft Defender for Cloud Apps was built to close that gap.

The gap today is the same shape, one layer higher.

Rogue Agents
Agents act — they don't just read

Unlike a shadow SaaS app, a rogue AI agent can send emails, modify SharePoint, and take autonomous actions 24/7 under a real user's delegated identity — with no IT ticket and no audit log by default.

Agent 365
Control plane is live — but not configured

Agent 365 provides a native inventory of every AI agent in your M365 tenant. Most organizations have never opened the management console or run an agent discovery scan.

Defender
Shadow AI discovery included in E5

Defender for Cloud Apps now surfaces unauthorized AI applications the same way it mapped shadow SaaS. The capability is bundled in M365 E5 — but AI app discovery policies aren't enabled by default.

With Copilot Wave 3 bundling agent creation into standard E3 and E5 licenses, there is no longer a procurement gate between an employee and a running autonomous agent. Any user with a Copilot Studio entitlement can build and deploy an agent in minutes. The question isn't whether agents are running in your environment. It's whether anyone is watching.


Why shadow AI carries more risk than shadow IT ever did

When an employee connected a personal Dropbox account to their work laptop, the blast radius was their files. When an employee builds or authorizes an AI agent through Copilot Studio, a third-party plugin, or a browser extension with AI capabilities, the blast radius is everything that user's Entra ID identity can reach — SharePoint, Exchange, Teams, OneDrive, and any connected line-of-business application.

Agents don't just read data. They take actions: drafting and sending emails, creating calendar events, modifying documents, querying APIs. They do this continuously, without human confirmation, at machine speed. An agent configured with overly broad permissions doesn't wait for an audit to cause harm — it acts between the moment it's deployed and the moment someone notices the anomaly.

Shadow IT Era Shadow AI Era
Discovery tool Defender for Cloud Apps Defender for Cloud Apps + Agent 365
What's hidden SaaS apps, OAuth tokens Autonomous agents, Copilot plugins, LLM workflows
Blast radius One user's files Delegated identity — SharePoint, Exchange, Teams
Action scope Passive (read, store) Active — sends, modifies, creates autonomously
Audit trail App access logs Purview AI Hub activity logs (if configured)
Remediation Block app, revoke OAuth Quarantine agent, restrict Entra ID scope, reconstruct audit log

Organizations running M365 Business Premium or E3 without E5 security add-ons are most exposed. Purview AI Hub's sensitivity label enforcement and Defender for Cloud Apps' AI application discovery both require E5 or equivalent licensing. E6 bundles the full stack — including Copilot licenses — and includes AI security posture management natively.


Three shadow AI gaps most organizations have right now

Gap 1 — No AI agent inventory: Security teams can list authorized SaaS applications. Almost none can answer: how many AI agents are active in this tenant, who built them, what identity they run under, and what data they can reach? Agent 365 provides this inventory natively — but it requires a deliberate configuration step that most organizations haven't taken. Without an inventory, you cannot govern what you cannot see.

Gap 2 — Inherited permissions, not least-privilege scope: An agent built in Copilot Studio inherits the Entra ID permissions of the user who created it. If that user has broad SharePoint access, so does the agent — and every action it takes, it takes as that person. Microsoft Purview sensitivity labels constrain what labeled content agents can reach, but only after the labels are deployed and enforced across the tenant. Most organizations actively rolling out Copilot have not completed this step, which means agents are already running against unlabeled data with no access boundary in place.

Gap 3 — No consumption guardrails or anomaly detection: Rogue agents don't just create compliance exposure — they run up Azure AI costs. An unauthorized or misconfigured agent running continuously can consume AI token budgets silently for weeks before anyone notices. Without hard token quotas in Azure Cost Management and Defender alerts on anomalous AI API call patterns, the first signal is an unexplained line item on the Azure invoice — after the activity has already occurred.


Seepath perspective

Microsoft built the full control stack for shadow AI. The gap is not in the product — it is in the configuration.

Agent 365 inventories every AI agent running in your tenant. Defender for Cloud Apps identifies unauthorized AI applications using the same discovery engine that mapped shadow SaaS. Purview AI Hub applies sensitivity labels to restrict what agents can access. Entra ID Conditional Access scopes agent identities to least-privilege boundaries. Azure Cost Management hard-caps token consumption so runaway agents don't drain budgets quietly. Every one of these controls is included in M365 E5 or E6. Most organizations with these licenses have not activated them for AI governance.

The playbook that contained shadow IT worked because organizations took a structured sequence: discover, classify, govern, monitor. Shadow AI requires the same sequence — but faster, because autonomous agents act at machine speed between detection events.

For organizations already deploying Copilot or building custom AI agents, the governance layer needs to be in place before broad rollout — not after the first incident. An AI Readiness Assessment maps your current agent footprint, identifies data exposure, and specifies the Agent 365, Purview, and Defender controls appropriate to your licensing tier. For organizations already on Managed AI Operations, shadow AI monitoring is part of the ongoing posture review cycle.

"Shadow AI isn't a future risk. If your users have Copilot and Copilot Studio access, autonomous agents are likely already running in your tenant. The question is not whether to govern them — it's whether the controls that came with your E5 license are actually turned on."

Assess your AI agent footprint →


Ready to Get Started?

Talk to a Seepath expert about your Azure, security, or AI strategy — no sales pressure.

Contact Us

Weekly Advisories

Get the latest Microsoft security, Azure, and AI updates delivered weekly.

Subscribe

Want Personalized Guidance?

Seepath has been a Microsoft Direct Bill CSP since 2014 — serving financial services and healthcare organizations.
with hands-on Azure, security, and AI implementation.

Free Azure Assessment Talk to an Expert