Skip to content
IT
Cybersecurity

Shadow AI: how SaaS updates hide invisible risks in companies

From the IBM Cost of a Data Breach Report 2026 to the silent accumulation of non-inventoried models

Published on By Sebastiano, GoYou.it 7 min read
Listen to the full article Full reading · 10:12 · from the channel GoYou Cybersecurity
Shadow AI: how SaaS updates hide invisible risks in companies

KEY TAKEAWAYS

  • Incidents related to shadow AI have doubled from 20% to 43% according to the IBM Cost of a Data Breach Report 2026
  • Shadow AI is not just about unauthorized chatbots: it includes models integrated into SaaS and internal applications through vendor updates
  • Most of these AI integrations happen transparently, through release notes ignored by security teams
  • A complete inventory of AI assets is essential to govern specific risks such as uncontrolled data flows to external models
  • Tools like ShadowAI-Watch can help identify undetected LLM connections in local processes

86% of companies ignore AI integrated into existing software

The 86% of organizations have not correctly inventoried the AI models integrated into existing software, according to the latest analysis by Thales Imperva. This visibility gap creates uncontrolled paths between sensitive data and language models, exposing companies to specific data breach risks and non-compliance with regulations such as the NIS2 Directive and the DORA Regulation.

Quick Answer

Shadow AI are not just unauthorized chatbots: they are language models integrated into existing software through vendor updates. The main problem is that 43% of security incidents involve non-inventoried AI, and most come through SaaS software updates. To mitigate the risk, it is necessary to first discover all the models active in the organization.

From chatbots to updates: the duality of Shadow AI

The concept of Shadow AI comprises two distinct but interconnected phenomena: the unauthorized use of chatbots by employees and the invisible integration of AI functionalities into existing software. IBM found that incidents related to the first phenomenon have doubled in a year, rising from 20% to 43%. However, the most insidious threat comes from the second category, where language models are introduced through vendor updates often ignored by security teams.

The illusion of compliance: when AI arrives via update

Most companies believe they have AI integrations under control, but reality is different. CRM, analytics platforms, and helpdesk systems have introduced advanced AI features such as virtual assistants, automatic summarization, and semantic search without security teams noticing. These changes arrive through release notes often overlooked, creating a visibility problem that goes far beyond simple unauthorized chatbots.

The specific risk of unmanaged AI integrations

Unmanaged AI integrations represent a specific and growing risk. Each non-inventoried model creates an uncontrolled path between sensitive data and third-party infrastructures. Unlike an unmanaged web application, which represents a forgotten door, an unmanaged AI integration is a forgotten door that can be manipulated through linguistic inputs. This creates compliance issues with regulations such as the NIS2 Directive and the DORA Regulation.

Why discovery is the new AI security priority

Before implementing policies or guardrails, organizations must first discover all active AI models. This discovery process has become the top priority for security teams, as highlighted by the Thales Imperva AI Application Security Framework. The strategy requires a complete inventory of all AI integrations, including models integrated into SaaS software through recent updates.

How to build a complete AI inventory

To address the Shadow AI problem, organizations must start with a complete census of all AI integrations. This includes customer-facing applications, internal assistants, developer tools, and CI/CD pipelines. For each integration, it is necessary to document which model is used, on which infrastructure it operates, which data it receives as input, which data it produces as output, and which team is responsible for it. Open-source tools like ShadowAI-Watch can help correlate local processes with external connections to language models, including integrations not passed through the procurement process.

The gap between perceived AI and real AI

Organizations often underestimate the extent of AI integrations present in their infrastructure. The gap between the AI that companies believe they use and the AI actually in use can be significant, often 100% more than initial estimates. This gap represents the true Shadow AI, introduced through vendor updates that go unnoticed in release notes.

Tools and resources for Shadow AI management

Thales Imperva offers a complete AI Discovery Playbook, which includes a worksheet for self-audit of the AI census. This tool helps organizations identify all AI integrations present in their infrastructure, assess the associated risks, and implement appropriate governance and protection measures. The playbook is available for download on the official Imperva website.

The importance of a structured approach to AI security

AI security requires a structured and methodical approach. Without accurate discovery and a complete inventory, organizations cannot implement effective policies or adequate protection measures. The Thales Imperva framework emphasizes the importance of discovering first, assessing risks, and then governing and protecting AI integrations. This systematic approach is essential to address the complexity and rapid evolution of threats related to the use of AI in corporate infrastructures.

Impact on compliance and risk management

The presence of unmanaged AI integrations can have a significant impact on compliance with regulations such as the NIS2 Directive and the DORA Regulation. Organizations must be able to demonstrate that they have complete control over all AI integrations to avoid sanctions and ensure operational resilience. Risk management related to AI integrations requires a combination of accurate discovery, risk assessment, and implementation of appropriate governance measures.

The challenge of managing AI integrations in cloud environments

AI integrations in cloud environments present a particular challenge for organizations. The dynamic and distributed nature of cloud infrastructures makes it difficult to maintain an accurate inventory of AI integrations. Tools for cloud security posture management (CSPM) and Security Information and Event Management (SIEM) can help identify and monitor AI integrations in cloud environments, but require accurate configuration and maintenance.

The importance of training and awareness

Managing AI integrations requires not only adequate tools and processes but also widespread training and awareness among development, operations, and security teams. Organizations must invest in employee training to ensure that everyone understands the risks associated with AI integrations and the best practices for managing them. Awareness and training are essential to prevent the emergence of new unmanaged AI integrations and to ensure that all integrations are adequately controlled and protected.

Integration with development and operations processes

Managing AI integrations must be integrated into existing development and operations processes. This includes the inclusion of AI security controls in continuous integration/continuous delivery (CI/CD) processes and the implementation of monitoring and incident response measures specific to AI integrations. Integration with existing processes ensures that AI integrations are managed consistently and that risks are mitigated proactively.

The need for a zero-trust-based approach

A zero trust architecture-based approach is essential for managing AI integrations. This approach requires that all AI integrations be authenticated, authorized, and continuously monitored. The zero-trust architecture ensures that only authorized AI integrations can access sensitive data and that all activities are monitored to detect abnormal or suspicious behaviors.

Managing AI integrations in hybrid environments

Organizations with hybrid environments, which combine on-premise and cloud infrastructures, must face particular challenges in managing AI integrations. The heterogeneous nature of hybrid environments requires a flexible and adaptable approach to managing AI integrations. Tools for identity access management (IAM) and disaster recovery as a service (DRaaS) can help ensure that AI integrations are managed consistently across all environments.

The importance of documentation and traceability

Documentation and traceability are essential for managing AI integrations. Organizations must maintain accurate documentation of all AI integrations, including the models used, the infrastructures on which they operate, and the data they access. Traceability ensures that all AI integrations are monitored and that any changes are documented and approved. Documentation and traceability are fundamental for compliance with regulations such as the NIS2 Directive and the DORA Regulation.

Integration with risk management processes

Managing AI integrations must be integrated into existing risk management processes. This includes identifying and assessing the risks associated with AI integrations and implementing appropriate mitigation measures. Integration with risk management processes ensures that AI integrations are managed consistently with other business risks and that resources are allocated efficiently.

Editorial Note and Disclaimer

The guides and content published on GoYou are the result of independent research and analysis activities, for informational, educational, and in-depth purposes.

GoYou does not constitute a journalistic publication or an editorial product pursuant to Law No. 62/2001 and does not provide real-time information.

The GoYou project does not provide professional, technical, legal, or financial advice and disclaims any liability for the improper use of the information published.

In the Crypto sector, every investment involves risks: readers are invited to always inform themselves autonomously before making any decision.

Source: imperva.com · GoYou.it is not a registered news outlet; all content is for informational and educational purposes.

Listen GoYou Cybersecurityevery article read in full
Go to the channel

You may also like

Behind every article is the work of our passion. If you enjoy GoYou.it, you can support its development and continuity.

Support with PayPal