Artificial intelligence is entering companies through many routes. While technology departments develop corporate projects, professionals bring tools into their daily tasks and some business areas deploy automations connected to internal applications. This adoption can bring agility, but it is also creating an increasingly significant problem: Shadow AI, or artificial intelligence used outside the organisation's control mechanisms.
The phenomenon has a growing technical dimension. According to Censys's State of the Internet 2026 report, AI-related infrastructure exposed on the Internet grew by 63 % in nine months. This figure does not mean that all those systems are vulnerable, but it does show an expanding technology surface that needs oversight.
The question is no longer only which tools employees use, but which AI systems are deployed, what data they can access and what actions they can carry out. For companies, regaining that visibility is the first step towards making the most of artificial intelligence without compromising security or losing control over their digital infrastructure.
What Shadow AI is and why it is a new business risk
Shadow AI is the use, development or deployment of artificial intelligence solutions without the authorisation, visibility or oversight established by an organisation. The concept comes from Shadow IT, the term used to describe applications and technology services that operate outside the technology department. AI, however, adds significant complexity: it can interpret information, connect to other systems and even carry out actions autonomously.
Until recently, the most common example was an employee who entered internal documentation into a generative AI application without knowing its privacy terms. Today, the problem can arise when a department connects a model to its CRM, adds an agent to a sales process or builds an automation with low-code tools without assessing its permissions, dependencies or security conditions.
The Cloud Security Alliance, in its research Shadow AI Infrastructure: The Invisible Enterprise Attack Surface, published in May 2026, warns that this phenomenon is evolving into an invisible enterprise infrastructure. It brings together autonomous agents, models connected to corporate data sources, automation workflows and systems developed without necessarily going through the usual technology assessment processes.
The main risk does not lie in an AI tool being external or having been brought in by a particular department. It appears when the organisation does not know it exists, cannot check how it uses information or lacks mechanisms to limit its behaviour. A seemingly simple integration can become a new way into critical documents, applications and processes.
That is why tackling Shadow AI requires broadening the traditional view of IT security. Having a list of approved applications is not enough. Companies also need to know which models they use, where they run, who administers them, how they connect to other systems and who is responsible for how they work.
The growth of AI infrastructure exposed on the Internet
The rapid expansion of enterprise artificial intelligence is also reflected in the infrastructure reachable from the Internet. Organisations increasingly use platforms to run models, develop intelligent applications and connect tools through application programming interfaces, known as APIs. These technologies enable new capabilities, but deploying them requires proper security configurations and ongoing management.
Censys's research on the state of the Internet in 2026 identifies an increase of more than 60 % in the public exposure of AI-related tools over nine months. The full 2026 report puts this growth at 63 %, a trend that shows how the technology surface associated with AI models, applications and services is widening.
The figure needs careful interpretation. A system being visible from the Internet does not necessarily mean it is vulnerable, lacks authentication or has been attacked. However, when that exposure coincides with outdated software, misconfigurations or excessive permissions, it can make unauthorised access easier and increase the impact of a security incident.
Among the technologies analysed by Censys are platforms related to running models and automation, such as Langflow and n8n. Their usefulness lies precisely in connecting components and services that previously worked independently. The risk arises when these connections are deployed without an inventory, maintenance or oversight, especially when credentials and business data sources are involved.
The consequence for organisations is clear: AI security cannot be limited to assessing the model being used. It must also cover the server it runs on, the APIs it receives, the tools it interacts with and the permissions that give access to corporate information. The infrastructure is part of the artificial intelligence system and, therefore, part of its risk surface too.
AI agents widen security and data access risks
The move towards artificial intelligence agents introduces a particularly significant change. While a conventional assistant answers the user's queries, an agent can combine instructions, consult different sources of information and use tools to complete tasks. This capability makes it possible to automate more complex processes, but it also requires precise control over which operations it can perform.
Imagine an agent that helps the sales department prepare proposals. To work, it consults internal documents, retrieves information from the CRM and uses an e-mail tool. If it has overly broad permissions or interprets malicious instructions contained in any of that content, it could attempt actions that go beyond its original purpose. The example illustrates a possible technical risk, not a specific incident attributed to any company.
The OWASP Top 10 for Agentic Applications 2026 project identifies threats specific to these systems, such as tool misuse, identity and privilege abuse, unexpected code execution and the propagation of errors between agents. These risks show why connecting a model to business applications requires controls beyond those used for an assistant that only generates text.
One of the fundamental aspects is applying the principle of least privilege. Each agent should have only the access it needs to perform its function, clearly distinguishing between reading, creating, modifying and deleting information. It is also important to limit which tools it can run, protect credentials and establish human oversight for particularly sensitive operations.
Another risk is the loss of traceability. If an automation modifies records, generates communications or interacts with several systems, the company must be able to reconstruct what happened and under what conditions. Without proper logs, an incorrect action can be hard to detect, investigate or correct. That is why the security of agents depends as much on the model as on the architecture of permissions, connections and oversight around it.
How to identify and control Shadow AI in a company
The response to Shadow AI should not be based solely on banning tools. Many professionals turn to artificial intelligence solutions because they need to reduce repetitive tasks, find information quickly or automate processes. If the company does not provide suitable alternatives, restrictions can encourage informal uses that are even harder to identify.
The first step is to build an artificial intelligence inventory that includes contracted applications, cloud services, private models, in-house developments, agents and automations. This inventory should also record who uses each system, for what purpose, what information it processes and which applications it is connected to. The aim is to uncover both the officially managed projects and the solutions brought in directly by teams.
The next step is to classify risks according to the sensitivity of the data and the operational capability of each system. A tool that helps draft public content does not need the same level of control as an agent able to consult financial information or modify customer records. This distinction makes it possible to set proportionate policies and focus the strictest controls on the uses that can have the greatest consequences.
The NIST AI Risk Management Framework proposes approaching risk management through four related functions: govern, map, measure and manage. This voluntary framework is useful for assigning responsibilities, assessing applications, documenting decisions and reviewing controls throughout the life cycle of systems. Security thus stops being seen as a one-off validation and becomes an ongoing activity.
In addition, the organisation needs to offer corporate environments where AI can be used in a controlled way. This means defining approved models, managing identities and permissions, logging relevant interactions and setting clear criteria for adding new integrations. The combination of the right tools, training and oversight makes it possible to reduce reliance on improvised solutions without holding back departments' useful initiatives.
Governance, regulatory compliance and technological sovereignty
Shadow AI also raises a question of corporate governance. When different departments use models, providers and automations independently, it becomes harder to apply common privacy, security and quality policies. The organisation can end up depending on fragmented infrastructure, with multiple credentials, contractual terms and responsibilities that are not always sufficiently defined.
In Europe, the European Union's Artificial Intelligence Act reinforces the need to know and assess the systems in use. It applies according to a staggered timetable, with some obligations already in force and others due at later dates, depending on the type of system. The existence of Shadow AI is not in itself an infringement, but a lack of inventory and oversight can make it harder to identify the obligations that apply to each organisation.
Technological sovereignty offers a strategic response to this complexity. It consists of keeping real decision-making power over the data, infrastructure, models and providers involved in business processes. It does not necessarily require using only in-house technology, but rather having an architecture that makes it possible to set the conditions of use, maintain traceability and avoid dependencies that compromise operational autonomy.
From this perspective, the security of artificial intelligence does not depend solely on where a model is hosted. It also requires controlling what information it receives, how its users authenticate, which applications it can consult and which logs make it possible to audit its activity. A private deployment without proper controls can have security problems, just as an external service can be part of a governed architecture if its access and data processing conditions are configured correctly.
At LurNova we advocate an approach to enterprise AI based on that ability to stay in control. Through solutions such as MIA, our corporate artificial intelligence platform, we make it easier to create assistants and agents that work with internal knowledge and allow users, permissions and traceability to be managed. For its part, LurNova AI Model Hub makes it possible to centralise the running, routing and monitoring of private and commercial models from an enterprise architecture.
Enterprise artificial intelligence needs visibility and control
Shadow AI reflects a reality that many organisations are beginning to face: the adoption of artificial intelligence can move faster than the mechanisms in place to manage it. Employees discover new tools, departments automate processes and technical teams add agents and models that extend the business's capabilities. The challenge is to make sure that growth happens within a known and supervised architecture.
The Censys data and the warnings from organisations such as the Cloud Security Alliance and OWASP show that AI security must cover much more than the content sent to a model. Access, credentials, integrations, infrastructure and the ability to carry out actions are also part of the risk. Without visibility over these elements, the company loses its ability to prevent incidents and to respond when something does not work as it should.
This does not mean giving up on innovation or subjecting every initiative to processes that prevent progress. It means establishing a technology foundation that makes it possible to experiment with, deploy and scale artificial intelligence solutions with clearly defined responsibilities. Inventorying systems, limiting permissions and keeping activity logs are measures that can improve security and support a more orderly adoption.
At LurNova we understand technological sovereignty as a company's ability to decide how it uses its artificial intelligence and to keep control over its digital assets. Because if artificial intelligence isn't secure, it isn't intelligent. And the first step towards ensuring that security is knowing which systems are running, what they connect to and who oversees them.
If your organisation is adopting agents, automations or AI tools, LurNova can help you assess their architecture and define a secure, traceable enterprise environment tailored to your needs.
Co-founder and CTO of LurNova, with more than 20 years' experience in developing digital technologies. Specialised in designing secure, private and governed artificial intelligence architectures and solutions for organisations.
