# Shadow AI: The Hidden Risk of Unsanctioned ChatGPT, Copilot and Gemini Use > Shadow AI, the unsanctioned use of public AI tools like ChatGPT, creates serious security risks, including data leakage, IP loss, and compliance breaches. Source: https://loopbackup.com/blog/shadow-ai-unsanctioned-chatgpt-copilot-gemini Publisher: Loop Backup Content language: en --- ## At a Glance * **What is Shadow AI?** The use of unapproved generative AI tools like ChatGPT, Gemini, or Copilot by employees, mirroring the long-standing problem of 'Shadow IT'. * **Why is it a risk?** Staff may paste confidential data, source code, customer PII, financial results, strategic plans, into public AI models, potentially exposing it to the world. * **Real-World Leaks:** In April 2023, Samsung staff accidentally leaked sensitive source code and internal meeting notes to ChatGPT on three separate occasions. * **The Copilot Paradox:** Even sanctioned tools like Microsoft 365 Copilot carry risks. 'Copilot oversharing' can expose sensitive data if your underlying SharePoint and OneDrive permissions are not properly configured. * **How to Mitigate the Risk:** A multi-layered approach is needed, combining technical controls like AI DLP, a clear generative AI policy, and crucial employee training. * **The Role of Backup:** In the event of a data leak, a robust backup is essential for forensic review to understand what was lost. It also provides a safety net when remediating permissions before a Copilot rollout. ### What is Shadow AI? For years, security teams have battled 'Shadow IT', the unsanctioned use of applications and services by employees outside of the IT department's purview. Now, a new, more potent variant has emerged: **Shadow AI**. This refers to the burgeoning use of public generative AI tools by employees without the organisation's explicit approval or oversight. Driven by a desire for increased productivity, employees are turning to consumer-grade AI assistants like OpenAI's ChatGPT, Google's Gemini, and others to help them write code, summarise documents, and draft emails. The problem is that these public platforms were not designed for enterprise use. They often lack the security, privacy, and compliance controls that businesses require, creating a significant, hidden risk of **ChatGPT data leakage** and intellectual property loss. ### The Lure of the AI Assistant It is crucial to understand that employees using these tools are rarely acting with malicious intent. They are simply trying to be more efficient. A developer might paste a code snippet into ChatGPT to find a bug; a marketing manager might upload a transcript to Gemini for a quick summary; a financial analyst might input raw data to generate a report draft. In their eyes, they are leveraging cutting-edge technology to perform their duties better and faster. They are likely unaware that many public AI models can use the data they receive to train future versions, meaning their 'private' input could become part of the public model, accessible to others. This fundamental misunderstanding is the primary driver behind the spread of Shadow AI. ### A Real-World Cautionary Tale: Samsung's 2023 Data Leaks The theoretical risk of Shadow AI became a stark reality in April 2023. It was widely reported that engineers at Samsung, one of the world's leading technology companies, had accidentally leaked sensitive internal data to ChatGPT. Within a period of just a few weeks, there were three separate incidents: 1. An engineer pasted faulty source code from a semiconductor database to identify a fix. 2. Another employee shared code from a proprietary program to optimise it. 3. A third staff member submitted the audio recording of an internal meeting to have ChatGPT generate the minutes. In each case, sensitive intellectual property was uploaded to a third-party platform outside of Samsung's control. The incident forced Samsung to ban the use of generative AI tools and begin developing its own internal alternative. It serves as a powerful warning to every organisation: if it can happen at Samsung, it can happen anywhere. ### Discovering the Shadows: How to Find Unsanctioned AI Use You cannot manage what you cannot see. Identifying the use of Shadow AI is the first step toward mitigating the risk. Several techniques can help shed light on the problem: * **Network and DNS Log Analysis:** Your firewall and DNS server logs are a valuable source of intelligence. By analysing logs for requests to domains like `chat.openai.com`, `gemini.google.com`, and other known AI services, you can build a picture of which tools are being accessed and by whom. * **Cloud Access Security Broker (CASB):** A CASB solution sits between your users and cloud services, providing visibility and control. Modern CASBs can identify traffic to thousands of applications, including generative AI tools, and can be configured to block or limit access based on your policies. * **Endpoint and Browser Telemetry:** Analysing browser history and network connections from corporate devices can also reveal the use of unsanctioned web-based AI tools. ### The Copilot Conundrum: When Sanctioned AI Is Still a Risk While Shadow AI deals with unapproved tools, a significant risk also exists with sanctioned, enterprise-grade AI like Microsoft 365 Copilot. The danger here is not the tool itself, but the data it has access to. The issue is **Copilot oversharing**. Copilot is designed to be deeply integrated into the Microsoft 365 ecosystem. It can reason over all the data a user has access to, their emails in Outlook, their chats in Teams, and their files in SharePoint and OneDrive. This is its greatest strength, but also its greatest potential weakness. Copilot meticulously respects existing permissions. The problem is that in most organisations, permissions are a mess. Over the years, through staff turnover and ad-hoc collaboration, SharePoint sites and Teams channels accumulate broken or overly permissive access rights. A sensitive document containing payroll information might have been accidentally shared with 'Everyone except external users' years ago. A user might not know this document exists, but Copilot does. If that user asks Copilot an adjacent question, the AI could draw upon and surface data from that sensitive file, instantly turning a latent permissions issue into an active data breach. Before rolling out Copilot, organisations must conduct a thorough audit and remediation of their Microsoft 365 permissions, a process that can be complex and time-consuming. ### Establishing Control: Policies and Technology Addressing the risks of both Shadow AI and Copilot oversharing requires a strategic, multi-pronged approach encompassing policy, technology, and people. #### 1. Draft a Clear Generative AI Policy Your first step should be to develop and communicate a formal **generative AI policy**. This document eliminates ambiguity and provides clear guidance for all employees. It should explicitly state: * An inventory of approved and prohibited AI tools. * A definitive list of data types that must **never** be entered into any public AI tool (e.g., customer data, PII, source code, financial information, health records, company strategy documents). * Guidelines for the responsible use of approved tools. * A requirement for employees to fact-check and verify any output from an AI, which can be prone to errors or 'hallucinations'. * Clear accountability for any work produced using AI assistance. #### 2. Implement AI-Aware Data Loss Prevention (AI DLP) Modern Data Loss Prevention (DLP) tools are becoming AI-aware. **AI DLP** solutions can go beyond simply blocking a domain. They can inspect the content being sent to web forms and browser-based applications. This allows for more granular controls, such as allowing employees to access ChatGPT but blocking them from pasting any content that matches a pattern for sensitive data, like a customer database or proprietary code. #### 3. Train Your People Technology and policies alone are not enough. You must educate your employees about the risks. Run training sessions that explain *why* pasting confidential information into public AI tools is dangerous. Use real-world examples like the Samsung incident to make the threat tangible. The goal is to create a culture of security awareness where employees become partners in defending the organisation's data. ### The Critical Role of Backup in Forensic Review and Remediation In this new landscape, a robust, immutable backup strategy becomes more critical than ever, particularly for incident response and proactive risk management. Should a **ChatGPT data leakage** event occur, your security team's first task is to conduct a forensic review to determine the exact scope of the breach. What specific data was lost? Which version of the source code was compromised? If the original data on the live system has been altered or deleted, a reliable, third-party backup is the only way to establish a ground truth. Services like [Loop's AI Cloud Backup](/ai-cloud-backup) and [Microsoft 365 backup](/microsoft-365-backup) provide an independent, immutable copy of your data, giving you the forensic fidelity needed to respond effectively. Furthermore, as you prepare for a safe M365 Copilot rollout, you will inevitably need to perform a large-scale clean-up of SharePoint and Teams permissions. This process carries its own risks. Having a point-in-time backup of your M365 data provides a vital safety net, allowing you to restore files and settings if the permissions remediation process goes wrong. ## Conclusion The rise of generative AI presents a transformative opportunity for productivity, but it also introduces profound new risks. Shadow AI is not a hypothetical threat; it is active in most organisations today, whether they know it or not. The convenience of these tools all but guarantees their use by well-meaning employees, creating a hidden pathway for your most sensitive data to leave the organisation. By combining proactive discovery, a clear and well-communicated **generative AI policy**, modern technical controls like **AI DLP**, and a robust backup strategy for forensic readiness, you can manage the risk. The goal is not to block progress, but to enable the safe and responsible adoption of AI, turning a hidden risk into a managed, strategic advantage.