Agentic AI and Shadow AI Are Creating a New Attack Surface Telecom Cannot Ignore

Dan O'Malley

Expert in 5G

Table of Contents

TL;DR

Shadow AI in telecom means AI tools or autonomous agents are being used inside operator environments without IT or security approval, visibility, or audit trail.
Agentic AI changes the risk because agents do not only generate output; they can use tools, access systems, call APIs, and take action.
AI agents should be treated as non-human identities because they can hold credentials and operate across systems like a user account.
AI agent drift creates permission creep when an agent gradually moves beyond its original approved scope.
MCP security risk matters because AI agents connected through tool protocols may gain broader access than intended if authorization is not controlled.
Policy and discovery tools are not enough because governance often identifies risk after access has already been granted.
Zero Trust security and network segmentation help limit an AI agent’s blast radius before governance catches up.
Telecom operators should treat shadow AI as a network architecture issue, not only an employee behavior issue.

What Is Shadow AI Security in Telecom?

Shadow AI security in telecom is the practice of identifying, controlling, and limiting the risk created when AI tools or autonomous agents are used inside a telecom operator’s environment without formal IT or security approval.

This includes AI assistants, internal automation agents, unsanctioned copilots, AI-enabled workflow tools, or autonomous agents connected to network data, engineering notes, customer records, tickets, configuration files, APIs, or operational systems.

At first, shadow AI may look like a familiar shadow IT problem. Employees use tools that help them move faster. Engineers may summarize configuration notes. Operations teams may analyze logs. Support teams may use AI to draft responses or search internal documents.

The difference is autonomy.

A traditional software tool waits for a user to act. An AI agent can act on behalf of the user. It can retrieve data, call APIs, make recommendations, trigger workflows, and sometimes execute actions. In a telecom network, that shift creates a new security concern because the agent may become part of the operational environment before security teams can properly see, classify, or govern it.

For telecom operators, this is not only a compliance issue. It is a network security issue.

Telecom networks support carrier connectivity, enterprise SLAs, private wireless services, edge applications, emergency communication, and customer-sensitive data flows. If an unmanaged AI agent touches those environments with unclear permissions, the operator now has a new attack surface.

Why Agentic AI Creates a New Telecom Attack Surface

Agentic AI moves AI from assistant to actor.

That matters because telecom environments are built around systems that already have layered trust relationships. Network operations, OSS/BSS workflows, service assurance, private 5G environments, mobile core systems, and edge services often depend on controlled access between different tools and domains.

An autonomous AI agent can sit inside that environment and interact with multiple layers at once. It may summarize alarms, query logs, inspect tickets, recommend changes, access customer context, or connect to workflow automation.

The security issue is not that AI exists. The issue is that AI agents may operate with access that was originally designed for humans, devices, or applications.

McKinsey’s Risk & Resilience¹ practice describes AI agents as “digital insiders” because they can operate inside systems with different levels of privilege and authority. McKinsey also reports that 80% of organizations have encountered risky AI agent behaviors, including improper data exposure and unauthorized system access.

For telecom operators, this risk is larger than ordinary enterprise exposure. A single access mistake can affect network availability, operational continuity, customer trust, and downstream enterprise services.

Key data points:

McKinsey¹ reports that 80% of organizations have encountered risky AI agent behaviors.
Deloitte’s 2026 State of AI in the Enterprise found that only 21% of organizations report mature agentic AI governance.
Telecom operators face a higher-risk environment because AI agents may interact with network operations, core systems, edge services, and customer-sensitive workflows.

Why AI Agents Are a New Class of Network Identity

AI agents should be treated as a new class of non-human identity.

They can hold credentials. They can inherit user permissions. They can access tools. They can store memory. They can connect to internal data sources. They can take action across workflows.

That makes them different from a normal application and different from a human user.

A human employee usually has a role, a manager, a department, a location, and a review cycle. A device has an owner, a certificate, a posture state, and a network segment. A service account has a technical purpose.

An AI agent may not fit cleanly into any of those categories.

It may begin as a narrow tool for summarizing tickets. Later, it may be connected to network documentation. Then it may be allowed to analyze configuration records. Then it may be linked to automation workflows. Each step may appear reasonable, but the combined result is an unmanaged identity with expanding reach.

This is why telecom security teams should not treat shadow AI only as a user policy issue. They should treat it as an identity governance gap.
If an AI agent can access operational systems, it needs identity controls. If it can reach multiple environments, it needs segmentation. If it can act across tools, it needs authorization limits. If it can retain memory, it needs monitoring and review.

This is where telecom operators should connect shadow AI security with Zero Trust security. Every actor, human or non-human, should earn access based on identity, context, policy, and necessity.

How AI Agent Drift Creates Security Gaps

AI agent drift is when an autonomous AI agent’s behavior gradually moves away from its originally approved scope.

This is one of the most important shadow AI risks for telecom environments because it does not always happen suddenly. It can happen through small changes.

An agent may first be approved to summarize network tickets. Then it is connected to a knowledge base. Then it gets access to historical incident data. Then it is linked to configuration notes. Then it begins recommending operational actions.

Nothing in that sequence may look dangerous by itself.

But over time, the agent’s effective permission set becomes wider than originally intended. That is permission creep. In a telecom environment, permission creep can create real network risk because agents may interact with operational data, sensitive workflows, or infrastructure logic.

Static permission reviews do not always catch this kind of drift.

Security teams may know what the agent was approved to do on day one. They may not know what the agent can do after several integrations, workflow changes, memory updates, or tool connections.

This is why shadow AI security needs continuous visibility and architectural containment. Discovery is useful, but discovery alone does not reduce blast radius. Policy is useful, but policy alone does not stop lateral movement.

Telecom operators need to assume that some agents will be over-permissioned before they are discovered.

Why MCP Security Risk Matters for Telecom Operators

The Model Context Protocol, or MCP, is becoming important because it helps AI systems connect with tools and data sources.

That can make AI agents more useful. It can also increase authorization risk.

When agents connect through tool protocols, they may gain access to systems, data, or workflows beyond the original intent. If authorization is too broad, an AI agent may not need to exploit a vulnerability. It may simply use access that was granted too casually.

OWASP’s³ Agentic Security Initiative highlights that agentic systems require serious threat modeling around tool use, authorization, memory, and agent-to-agent communication.

For telecom operators, this matters because network environments are interconnected by design. Operations platforms, core systems, RAN systems, private 5G services, edge workloads, and customer-facing systems may depend on many layers of controlled access.

If an AI agent enters that environment through a tool connector, operators need clear answers:

Which agent is using the tool?
Which identity is attached to the agent?
Which systems can it access?
Can it communicate with other agents?
Can it retain sensitive memory?
Can it trigger actions or only produce recommendations?
Can it cross from a low-risk workflow into a high-risk operational domain?

The answers should not live only in an AI policy document. They should be enforced in the network.

How Zero Trust and Segmentation Reduce Shadow AI Risk

Shadow AI cannot be solved only by banning tools.

That approach is unrealistic because AI is already becoming part of telecom operations, engineering productivity, service assurance, customer support, and network planning. The better approach is to make AI use governable, observable, and contained.

This is where Zero Trust and segmentation become practical.

The same segmentation that limits lateral movement from a compromised endpoint can also constrain an over-permissioned AI agent. If an agent is misconfigured, compromised, or allowed to drift, segmentation can stop it from moving freely across unrelated systems.

This is why network segmentation vs. segregation is directly relevant to agentic AI risk. The goal is not only to organize traffic. The goal is to reduce blast radius.

For telecom operators, segmentation should apply across network domains, management systems, private wireless environments, edge platforms, and operational workflows. If an AI agent is approved for one narrow function, it should not automatically gain access to unrelated systems.

Network visibility also matters. Tools like Nexus NMS support operational awareness across telecom environments, which becomes more important as AI agents begin to interact with network workflows and infrastructure states.

Security enforcement should also sit close to the architecture itself. Nybsys security solutions such as Aegis Security Gateway are relevant to the broader security model because telecom operators need carrier-grade enforcement across trusted and untrusted network paths.

The goal is not to position shadow AI as one more tool category. The goal is to recognize the architectural reality:

An unmanaged AI agent is an unmanaged network actor.

If it can access systems, it must be identified. If it can take action, it must be governed. If it can move across domains, it must be segmented. If it can retain context, it must be monitored.

This applies across modern telecom environments, including private wireless, edge infrastructure, and 5G Core systems where access boundaries are critical.

Ultimately, the operators who get ahead of shadow AI will not be the ones who ban AI tools. They will be the ones whose network architecture already assumes every actor, human or otherwise, has to earn its access.

Explore More on Telecom Security & Networking

Frequently Asked Questions

What is shadow AI in telecom security?

Shadow AI in telecom security is the use of AI tools or autonomous agents inside a telecom operator’s environment without IT or security approval, visibility, or audit trail. It becomes a network risk when those tools interact with operational systems, customer data, credentials, network documentation, APIs, or automation workflows.

Why is agentic AI a security risk for telecom operators?

Agentic AI is a security risk because autonomous agents can use tools, access systems, call APIs, retain memory, and take action across workflows. In telecom environments, this can affect network operations, private wireless services, edge systems, service assurance, and core infrastructure if access is not properly governed.

Why are AI agents considered non-human identities?

AI agents are considered non-human identities because they can hold credentials and perform actions inside systems like a user account. Unlike traditional human users or devices, autonomous agents may change behavior over time, connect to new tools, and operate across workflows without fitting neatly into existing identity governance models.

What is AI agent drift?

AI agent drift is when an autonomous AI agent gradually moves beyond its originally approved scope. This can happen when the agent receives new tool access, connects to more data sources, retains more context, or begins supporting workflows it was not originally designed to handle.

What is MCP security risk?

MCP security risk refers to authorization and access-control concerns that arise when AI agents connect to tools and data sources through the Model Context Protocol. If access is too broad, an agent may reach systems or information beyond what was intended.

How does network segmentation help reduce shadow AI risk?

Network segmentation reduces shadow AI risk by limiting how far an AI agent can move if it is over-permissioned, misconfigured, or compromised. The same segmentation principles used to stop lateral movement from compromised devices can also limit the blast radius of autonomous AI agents.

Should telecom operators ban AI tools to stop shadow AI?

Banning AI tools is unlikely to work because AI is already becoming part of telecom operations and engineering workflows. A stronger approach is to make AI use visible, governed, identity-aware, and contained through zero trust, segmentation, access control, and continuous monitoring.