FluidOne Blog

Defending the unseen: How security teams can manage Shadow AI risk

Written by Nathan Long | 17/11/2026, 11:56

AI adoption rarely waits for an organisation’s governance programme to catch up. Employees can access powerful generative AI services in seconds, often without procurement, technical integration or any direct involvement from the security team.

That creates an understandable concern about “Shadow AI”. However, the presence of an unapproved AI tool does not automatically mean sensitive information has been exposed or a serious security incident has occurred. An employee using AI to improve the wording of a generic email does not present the same risk as someone uploading customer data, proprietary source code or an internal incident report.

For security teams, the challenge is not to eliminate every unsanctioned use of AI. It is to build enough visibility to understand how AI is being used, identify activity that could cause meaningful harm and apply controls proportionate to the information, user and use case involved.

Why are people using Shadow AI in the first place?

Before trying to control Shadow AI, I think it is important to understand why people are using it. In most cases, someone is not deliberately trying to bypass security policy. They have simply found a tool that helps them get something done quicker or better.

  • Summarise documents

  • Improve emails and reports

  • Analyse data

  • Generate or review code

  • Research unfamiliar subjects

  • Automate repetitive work

From a security perspective, the instinct to block unapproved AI services is completely understandable. As with most professionals in security, I would rather people did not use Shadow AI at all. But realistically, if somebody can access these tools, including outside the organisation’s technical controls, simply telling them not to use them does not solve the problem.

Some uses may be relatively low risk. Others could expose customer information, intellectual property, security data or commercially sensitive information. Distinguishing between the two is where the real challenge sits. Visibility is only useful if it provides context, rather than simply producing a blanket list of AI domains to block.

Not all Shadow AI represents the same risk

There is a major difference between asking an AI tool to improve a generic paragraph and uploading a report containing personal or sensitive information. In practice, the level of risk depends on several factors working together: the data involved, the service being used, the user’s role and how the output will be applied.

An enterprise-approved AI service with contractual protections, corporate identity integration and administrative controls should not be assessed in the same way as an anonymous consumer service or an unreviewed browser extension.

The user’s role also changes the potential impact. Someone working with public marketing content presents a different exposure from an administrator, developer, legal adviser, finance employee or incident responder with access to privileged information.

Finally, generating ideas is different from allowing an AI system to make, or materially influence, decisions affecting customers, employees, security operations or regulated processes. Taken together, those factors help an organisation distinguish low-risk experimentation from activity that needs stronger controls or direct intervention.

Visibility must extend beyond network traffic

Traditional web controls may show that a user accessed an AI service, but they will not always show what information was submitted, how the service was used or whether the activity was authorised. Security teams therefore need to combine several sources of visibility:

  • Secure web gateway and DNS telemetry

  • Endpoint and browser activity

  • Cloud application discovery

  • Identity and authentication logs

  • Data loss prevention controls

  • SaaS security monitoring

  • Application inventories and procurement records

  • User reports and engagement with the business

That visibility should be used carefully. The aim is not to monitor every employee interaction indiscriminately. It is to identify patterns of use, understand where sensitive information may be leaving controlled environments and prioritise investigation around meaningful risk.

When we have looked at implementing this type of detection within the SOC, one of the biggest considerations has been context. Simply knowing that somebody accessed an AI service is not particularly useful on its own. It can quickly generate a lot of noise without identifying the activity we actually care about preventing.

For example, a detection that says “a user accessed an AI website” has limited operational value. Knowing that a privileged user uploaded a document labelled confidential to an unmanaged generative AI service gives the SOC something much more meaningful to assess.

Data classification makes controls proportionate

Many organisations will be tempted to begin with a list of permitted and blocked AI services. That may provide an immediate layer of protection, but it does not answer the underlying question of which information can safely be used with AI.

A more mature approach connects AI controls to the organisation’s existing data-classification model:

  • Public information: AI use may be acceptable with minimal restrictions.

  • Internal information: Use may be permitted through approved services.

  • Confidential information: Additional technical, contractual and access controls may be required.

  • Restricted or regulated information: Submission may be prohibited unless a specifically authorised use case and platform have been approved.

This moves the organisation away from treating AI as a single category of risk. It also makes policy easier for employees to understand because the decision is based on the information they already handle, rather than expecting them to evaluate the technical architecture of every AI product.

Identity and access still matter

AI may feel like a new category of technology, but the same security principles still apply. Approved services should use corporate identities, multifactor authentication and appropriate conditional access controls. Organisations should also consider restricting personal accounts for business activity, controlling plug-ins and third-party integrations, and applying least privilege to the data an AI service can reach.

This becomes more important as AI tools move beyond simple chat interfaces and connect directly to email, documents, code repositories, ticketing platforms and other business systems. At that point, the risk is no longer limited to what somebody manually enters into a prompt. It includes everything the service or agent has permission to access.

Policy alone will not prevent Shadow AI

A policy that simply says “do not use unapproved AI” may protect the organisation’s position, but it is unlikely to stop employees using tools that help them work faster. Effective governance needs policy, education and technical controls to work together.

Policy: Give employees clear rules about acceptable use, prohibited information and when formal approval is required.

Education: Use realistic examples rather than abstract warnings. Employees need to understand why entering customer data, source code, credentials or internal reports into an unapproved service could create risk.

Technical controls: Use blocking, warnings, monitoring, data loss prevention and access controls where the potential impact justifies them.

If approved alternatives are slow, unavailable or difficult to use, people will continue looking elsewhere. Security teams therefore need to work with IT, legal, privacy, procurement and business leaders to provide safe and usable routes to adoption. Understanding where Shadow AI is being used may also show where the business has a genuine capability gap that needs to be addressed.

The takeaway

For me, Shadow AI comes back to something security teams already deal with every day: visibility without context only gets you so far. Blocking everything might reduce one type of risk, but it can also push behaviour somewhere the organisation has even less visibility.
The better position is to understand which services people are using, what information is involved, who is using it and what they are trying to achieve. From there, security teams can make a meaningful decision about the risk and apply controls proportionately.