# AI Agents Are the New Shadow IT—and Permission Is the Real Problem

> Why autonomous AI deployments should be governed by permissions and transactions, not clever prompts.

- Author: Maged Zamzam (https://magedzamzam.ae/profile)
- Published: 2026-10-13
- Topics: AI agents, Cybersecurity, Governance, Prompt injection, Opinion
- Canonical: https://magedzamzam.ae/notes/ai-agents-are-the-new-shadow-it

---
*We are giving probabilistic software access to deterministic systems and calling it productivity.*

An employee connects an AI agent to mail, chat, a project tracker and a cloud console. The agent summarizes a discussion, drafts a response and opens a change request. Everyone is impressed. Then the workflow grows: it can send, delete, approve, transfer, deploy and provision. At that point, the question is no longer whether the model is intelligent. It is whether the organization understands who—or what—has authority to act.

Here is the controversial part: **most enterprises should not deploy general-purpose autonomous agents with broad write access.** Not until they can explain their permission model and demonstrate transactional controls with the same rigor used for production applications.

## The old shadow IT problem has acquired an operator

Shadow IT traditionally meant unapproved software or unmanaged data flows. An agent adds an action layer. It can read instructions inside documents, interpret context, choose tools and execute sequences. A malicious paragraph hidden in a retrieved page can attempt to redirect a workflow; a confused model can choose an unintended tool even with no attacker present.

This is not theoretical taxonomy. OWASP's 2025 LLM risk guidance addresses prompt injection, sensitive information disclosure and excessive agency. Excessive agency arises when an AI-enabled application has capabilities or permissions beyond what is necessary for the task, creating the possibility of harmful actions. [1][2]

Calling this a prompt-engineering problem is too optimistic. A stronger system prompt is not a security boundary.

## An AI agent is an untrusted planner with tools

Treat the model as a planner whose outputs are suggestions until independent software authorizes them. A safe workflow separates four roles:

1. **Identity:** a human or service principal with a defined scope.
2. **Planning:** the model proposes an action with structured arguments.
3. **Policy enforcement:** deterministic code checks permissions, resource scope, limits and preconditions.
4. **Execution and evidence:** a narrowly scoped tool performs the action and emits an auditable result.

Do not give the model raw administrator credentials. Do not assume a human approval button compensates for a vague or misleading action summary. The approval must show the actual target, parameters and consequence.

## A simple thought experiment

An assistant is asked to “clean up old customer files.” Does that mean archive duplicate exports, remove temporary files, delete storage buckets, or terminate retention under a legal hold? A language model can construct plausible answers to each interpretation. The system of record must define what is permitted, not the model's prose.

A safer flow generates a proposed change set, checks classification and retention policies, then requires approval for destructive changes. The executor receives an allowlisted resource set and an expiring capability token, not standing access to every drive.

For transactions with material financial or operational consequences, I favor two-person review or independent control points. This may feel less magical than an autonomous demo. That is precisely the point: production systems are not judged by the beauty of a demo.

## The controls I would insist on

| Layer | Minimum control |
|---|---|
| Tool access | Least privilege, scoped credentials and explicit allowlists |
| Inputs | Label externally retrieved content as untrusted instructions |
| Execution | Idempotency keys, parameter validation and transaction limits |
| Approval | Human confirmation of the *actual* write action when high impact |
| Monitoring | Action logs, correlation IDs, anomaly detection and revocation |
| Evaluation | Adversarial tests for injection, unauthorized requests and ambiguity |

NIST's generative AI risk-management profile offers a complementary governance lens: organizations need to assess and manage risks across the system lifecycle, not only at deployment. [3]

## Is the argument too conservative?

There is a credible counterargument. Human operators make mistakes, toil is expensive, and narrowly designed agents can execute bounded tasks more consistently than ad hoc scripts. I agree. Agents are valuable precisely when they remove repetitive coordination from a well-understood workflow.

My objection is to conflating *automation of a bounded task* with *delegation of broad authority*. A read-only diagnostic agent is a different risk class from an agent authorized to delete infrastructure. A purchasing assistant preparing a cart is not equivalent to one releasing an unrestricted payment.

Organizations should graduate autonomy by evidence: start with read-only visibility, then draft proposals, then bounded reversible actions, and only later consider privileged operations. Measure wrong-action rate and policy-violation attempts, not just time saved.

## The governance question is a product question

If a vendor cannot explain exactly how agent permissions are scoped, how hostile retrieved text is isolated, and how an erroneous action is reversed, that uncertainty belongs in the purchase decision. Procurement questionnaires should ask to see the executor's authorization path, not merely the model's safety promises.

The future of AI agents may be enormous. But **the trustworthy agent will not be the one that can do anything. It will be the one that can prove why it was allowed to do exactly this thing—and nothing more.**

![Conceptual diagram: AI Agents Are the New Shadow IT—and Permission Is the Real Problem](https://magedzamzam.ae/media/c65fe7fd3fc28dd3bb30.jpg "Conceptual design, not a deployment blueprint")

### Sources

[1] [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/)

[2] [OWASP LLM06: Excessive Agency](https://owasp.github.io/www-project-top-10-for-large-language-model-applications/2_0_vulns/LLM06_ExcessiveAgency.html)

[3] [NIST Generative AI Profile, AI 600-1](https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence)

---

Source: https://magedzamzam.ae/notes/ai-agents-are-the-new-shadow-it. Views are the author's own.
