How to Stop Shadow AI Without Banning AI Agents

Shadow AI is not a detection problem: no platform, including Atipol, finds every unmanaged agent on its own, so the fix has to make the declared path faster than the undeclared one. That means a catalog of approved MCP servers, permissions set tool by tool, and administrator-logged exceptions for anything outside the catalog, applied the same way to every declared agent.
Banning AI agents does not survive contact with a deadline
A DSI who issues a blanket ban on connecting AI agents to company systems is solving the easy half of the problem. Banning does not stop a team under deadline pressure from wiring an agent into a spreadsheet, a ticket queue, or a customer database over a weekend, the same way banning unapproved SaaS tools never stopped shadow IT; it only guarantees that when it happens, nobody told the DSI first. The pain is not that agents exist outside the plan, it is that there is no framework to hand a team the moment they ask how to do this properly instead of quietly.
The instinct that follows a ban failing is to look for a tool that finds every unmanaged agent and MCP server already running, the network-scanning equivalent of shadow IT discovery. That instinct runs into a harder limit than policy: an agent is a process a developer can start from a laptop with a single command, often talking to a service over a protocol built to be lightweight and easy to stand up, which is a much smaller footprint to hide than a SaaS subscription with an invoice trail.
Treating every agent the same way, whether it drafts a status update or touches customer records, wastes the same effort a blanket ban wastes: it either blocks harmless automation outright, pushing it underground, or it waves everything through because enforcing a ban case by case is too slow, which leaves the agents that actually carry risk no more controlled than before the policy existed.
No platform detects shadow AI on its own, including Atipol
Atipol is explicit about this limit rather than quiet about it: it does not, on its own, detect MCP servers or agents installed outside the platform, and does not replace a separate discovery process for unmanaged tools. An agent that was never declared to Atipol, and never connects through it, leaves no event for Atipol to log, the same way a server a security team never onboarded leaves no event in that team's own monitoring. No platform, Atipol included, closes that gap by scanning for what was never pointed at it in the first place.
This is not a reason to give up on governance, it is a reason to put the energy into the part of the problem a framework actually solves: making the declared, approved path fast and clearly better than the undeclared one, so fewer teams have a reason to go around it, and giving the teams that already have agents running a specific, low-friction way to bring them into view rather than leaving that decision to a ban nobody expects them to follow.
OWASP's own guidance on Excessive Agency makes the same point from a different angle: reducing what an agent can do, minimizing its functionality, permissions, and autonomy, closes more of the gap a detection tool cannot, because a narrowly scoped agent has less to abuse even if it is never found. NIST's AI Risk Management Framework frames the same idea at the level of an entire organization, as a voluntary framework meant to help build trustworthiness considerations into the design and development of a system, not to scan for what was never declared to it after the fact.
In one sentence:
A framework does not find an undeclared agent; it makes the declared path fast enough, and the catalog visible enough, that fewer teams have a reason to skip it. No platform, including Atipol, closes the gap by scanning for what was never pointed at it.
What a framework offers instead of detection
The shape of that framework is a catalog rather than a scanner: Atipol relies on a catalog of servers allowed by the organization, set to one of three modes, open workspaces, catalog only, or mandatory approval, with the rule that a workspace can tighten the rules, never loosen them. Each person gets their own agent lead, which acts for that person and never has more rights than they do, so an agent's identity in the system is tied to a declared human from the moment it is created, rather than to an anonymous script nobody can trace back to a request.
The catalog does not need to start by listing every tool a team might eventually want; it starts by listing what is already approved, with everything else routed to a specific, visible decision. "Off-catalog requests are decided by an administrator, and logged," which turns the question of a new, unlisted MCP server from a silent connection into a recorded request an administrator either grants or refuses.
Put side by side, the two paths look like this:
| Undeclared agent | Declared agent (catalog) | |
|---|---|---|
| Visible to an administrator | No record exists | Logged at declaration |
| Tool permissions | Whatever the process can reach | Set tool by tool |
| Off-catalog request | No request to make | Decided and logged by an administrator |
| Identity | Anonymous script | Tied to one person's agent lead |
Shadow AI risk against the framework's response
The Model Context Protocol's own security guidance names the specific risks an undeclared or poorly scoped agent connection creates; a declared catalog with tool-by-tool permissions answers each one directly rather than by trying to detect the connection itself.
| Risk (MCP security guidance) | What it looks like unmanaged | Framework response |
|---|---|---|
| Token passthrough | A server forwards a client's token to a downstream API without validating it was issued for that server | A catalog entry ties a server to a specific, reviewed configuration before any agent can use it |
| Scope inflation | A token is granted broad scopes such as files:* or admin:* because that was easier than asking twice | Permissions are set tool by tool, allowed, approval, or denied, not as one blanket grant per agent |
| Local server compromise | A locally run server executes with the same privileges as whoever started it, unsandboxed | Off-catalog servers are routed to an administrator decision before they run at all |
| Confused deputy, consent skipped | A proxy server reuses a prior consent cookie to skip a new client's authorization screen | Every declared agent is tied to one person's agent lead, which never exceeds that person's own rights |
| Undeclared connection | Nobody outside the team that started it knows the agent or server exists | A catalog-only or mandatory-approval workspace makes an undeclared connection fail to start rather than fail to be noticed |
Build the catalog around least privilege, not blanket trust
The Model Context Protocol's own security best practices describe the alternative to blanket trust as a progressive scope model: a "minimal initial scope set containing only low-risk discovery/read operations," with "incremental elevation via targeted... scope challenges when privileged operations are first attempted," rather than a server publishing, or a client requesting, every scope it might ever need up front. The specification is direct about the failure mode this avoids: "publishing all possible scopes," or "using wildcard or omnibus scopes (*, all, full-access)," expands what a single stolen token can reach and makes it harder to revoke access without disrupting every workflow that token touched.
The same document's guidance on locally run servers applies the identical principle to the machine itself rather than to a token: a client should "execute MCP server commands in a sandboxed environment with minimal default privileges," launched "with restricted access to the file system, network, and other system resources," with explicit consent shown before a new local server runs at all. A catalog built around these two ideas, minimal scope by default and sandboxed local execution, gives a declared agent less to abuse even when a review misses something, which a detection tool aimed at the undeclared agent cannot do.
The same guidance is specific about what that consent step has to show before a local server runs at all: the exact command that will be executed, without truncation, a clear identification that it is a potentially dangerous operation that executes code on the user's system, and explicit approval required before the configuration proceeds. Highlighting command patterns that touch a home directory, SSH keys, or system directories, and warning that a server runs with the same privileges as the client itself, turns a one-click install into a visible decision a person makes rather than something that happens silently in the background.
Make the exception path visible instead of invisible
An administrator who decides and logs every off-catalog request is not a detection system, but it produces something a detection system would also need: a record of who asked for what tool, when, and whether it was granted. That record exists because the decision was made inside the declared path, by a person routing a request through the catalog, not because a scanner found a connection after the fact. The honest version of this framework tells a team exactly that distinction up front, rather than letting them assume a catalog quietly watches for what they connect outside it.
This also sets the right expectation for what happens when an undeclared agent is found some other way, through a vendor invoice, a network alert unrelated to Atipol, or a colleague mentioning it in a meeting. The catalog gives that agent somewhere to go once it is found: a category to request, a workspace mode that already applies to it, and an agent lead to assign it to, rather than a policy decision made from scratch each time one surfaces.
This also keeps the exception auditable in the same vocabulary the Model Context Protocol's own guidance recommends for scope elevation: a specific, logged request met with a specific, logged decision, rather than a client silently assuming broader access because asking twice felt like friction.
Giving the framework to the teams already running agents
The three workspace modes move in one direction only:
A declared-by-default rollout works better than a ban followed by a hunt. Moving a workspace from open to catalog only, then from catalog only to mandatory approval, matches the three modes the platform already supports, and matches the principle that rules only tighten, never loosen, so a team never experiences the framework as a surprise step backward in what they can already do. Each agent gets its own agent lead at declaration time, which ties identity and permission to a specific person from day one, the same structural choice the Model Context Protocol's own guidance makes when it recommends binding a session to a specific user rather than treating it as an anonymous credential.
A short grace period for declaring agents already in use, paired with the same catalog and permission rules applied going forward, converts a backlog of undeclared automation into inventory without requiring anyone to admit to having broken a rule that, in practice, nobody could have followed in the first place.
None of this requires finding every agent already running before starting; it requires making the declared path the one teams prefer, backed by the permission model, the catalog, and the logged exceptions described above, at atipol.ai/agentic-ai-blog/mcp-security and atipol.ai/agentic-ai-blog/ai-agent-governance, and the guided tour of that model at atipol.ai/security.
Frequently asked questions
Can a platform detect shadow AI agents automatically?
No platform does this reliably, including Atipol: it does not, on its own, detect MCP servers or agents installed outside the platform, and does not replace a separate discovery process for unmanaged tools. An agent that never connects through a platform leaves no event for that platform to log, which is why a catalog and permission model, rather than a detection claim, is the realistic response to shadow AI.
What should a catalog of approved MCP servers actually control?
A catalog controls which servers an organization allows at all, set to one of three modes, open, catalog only, or mandatory approval, with permissions applied tool by tool rather than as one blanket grant per agent. The Model Context Protocol's own guidance adds that each server should start with a minimal scope and only request more through a specific, logged elevation, not a wildcard grant up front.
What happens when someone requests a tool outside the approved catalog?
The request is decided and logged by an administrator rather than silently refused or silently allowed. That decision becomes part of the record precisely because it went through the declared path, which is the mechanism a framework offers in place of trying to detect a connection that bypassed the catalog entirely.
Sources
- Model Context Protocol, Security Best Practices, version 2025-11-25, consultée le 11/11/2026
- OWASP GenAI Security Project, LLM06:2025 Excessive Agency, consultée le 11/11/2026
- NIST, AI Risk Management Framework (AI RMF 1.0), publié le 26/01/2023
The public demo shows a fictional organization at work, no account needed. The free plan includes a project, three agents and an agent lead.
