Shadow IT in German Businesses: The NENNA Guide to Cybersecurity Risks

Shadow AI in German Businesses: The NENNA Guide to Cybersecurity Risks

Every day, confidential information leaves businesses completely unchecked. This is how business-critical data ends up in AI systems operating outside of European jurisdiction. It often happens unconsciously, frequently because no alternative is available or because organisations are simply unaware of the actual risks.

Our NENNA Guide to Shadow AI Usage:

What is Shadow AI and What Makes It So Dangerous?

Shadow AI refers to any AI application used within an organisation without the knowledge or approval of the IT department. It arises when employees use personal accounts or free browser-based access to ChatGPT, Gemini, DeepSeek, or other tools to write emails, summarise documents, generate social media graphics, or create presentation slides – all without internal IT clearance. In practice, this means:

According to the global KPMG study "Trust, attitudes and use of AI: A global study 2025" (KPMG & University of Melbourne, 2025), 73% of employees are already using generative AI in the workplace, mostly via freely accessible tools. According to Bitkom e.V., 42% of German companies are already aware that their employees are using private AI tools like ChatGPT or Gemini for business purposes. Another 17% suspect it but cannot prove it with certainty. This share has doubled compared to the previous year. What makes this particularly dangerous for businesses: almost one in four companies do not know the actual situation in their organisation. This means control over sensitive data is slipping away from many organisations faster than processes can be adapted. (bitkom, 2025)

The Uncalculable Data Protection Traps

Inputs into such tools – such as prompts or uploaded documents – are stored and used for AI training. Many public AI tools store inputs on servers in non-European jurisdictions, including the USA or China. Gemini, Google's AI service, stores entered prompts in its free browser version for 18 months on servers in the United States, unless you manually turn this off or change the retention period. Even if you have taken the trouble to disable the "Save Activity" setting, Google still uses chats to reply to you and keep Gemini safe, partly by using human reviewers. DeepSeek stores all prompts directly in China: "When you use our Services, we may collect your text input, prompt, uploaded files, feedback, chat history, or other content that you provide to our model and Services".

As a result, trade secrets become part of the AI's knowledge base. In generative AI with web search enabled, entered company details are systematically processed and permanently embedded in the model. In some cases, web queries that are pre-activated are sent to search engines like Google and stored there as well. This means that once data is entered, it is out of your control from that exact moment. The consequences for confidentiality, data protection, and compliance are difficult to assess. Corporate or customer data entered today could appear in answers for your competitors tomorrow.

According to the international KPMG study, 48-49% of respondents admitted to uploading sensitive corporate data to public AI systems – including financial figures, sales and customer data, or even copyrighted material. Paradoxically, such violations happen most frequently where companies have introduced rules. 67% of employees in companies with an AI ban and 56% in companies with an AI policy still upload sensitive data to public tools. In organisations with no policy (33%) or an unclear framework (38%), this happens much less frequently. (KPMG & University of Melbourne, 2025)

Despite Concerns About Data Loss: Many Companies Have No Guidelines

85% of AI users have concrete cybersecurity concerns. Nevertheless, they use AI tools because they are easily accessible and simple to use. (KPMG & University of Melbourne, 2025) These low barriers mean that shadow AI is also being deployed in sensitive business processes, prioritising convenience over security.

Only 23% of companies have formulated their own rules for using AI so far – and even these often do not cover all shadow applications (bitkom, 2025). The figures clearly show that bans do not work, and policies alone do not guarantee security. Without concrete guidance, clear processes, and real training on responsible AI usage, shadow AI risks will emerge – regardless of what is officially permitted or forbidden. (KPMG & University of Melbourne, 2025)

Furthermore, studies show that companies lack a complete overview of where AI is being used. The attack surface is growing faster through AI systems than governance structures can keep up. (Deloitte, 2025) Security assessments are missing before new AI tools are deployed (World Economic Forum, 2025b). According to Tagesspiegel Background, reporting on a Microsoft study, the situation is even clearer: in politics and administration, nearly one in two people (45%) use unverified AI tools despite known risks. At the same time, the public increasingly views AI as a security threat: 67% fear the misuse of AI for cyberattacks, and 66% fear the manipulation of AI systems. Yet, many fail to act: only 43% take any protective measures at all. This is compounded by a massive knowledge deficit: 73% feel poorly informed about AI, and 69% worry about the protection of their personal data (Tagesspiegel, 2025).

This mixture gives rise to shadow AI with numerous risks: data leaks, model manipulation, compliance violations, and liability risks can be the result. Businesses are inadvertently creating their own vulnerability.

While professional attackers have long been using AI for deepfakes, phishing, or automated malware, unverified services are frequently deployed in corporate environments – increasing their own attack surface through prompt leaks or accidental data exposure.

Technical Dead End: Why Classic IT Security Is Not a Comprehensive Fix

Conventional firewalls or monitoring solutions struggle to detect the use of AI services because AI traffic looks like normal web traffic. Business-critical information can slip into external AI models unnoticed, without traditional protection mechanisms raising the alarm. The only exception is if the use of AI is completely blocked in the browser. But even then, AI is increasingly integrated into existing tools and is almost impossible to keep track of. Monitoring these constant developments would require immense effort. Furthermore, there is always the risk that employees will simply use their personal devices to access AI.

Companies are losing track of what data is leaving the organisation and how it is being processed. This creates significant compliance risks, liability issues, and potential reputational damage if confidential information becomes known to third parties via AI models.

Recommendations for Businesses

In practice, this means: if you want to curb shadow AI, you must establish clear rules for the use of Artificial Intelligence. Bitkom e.V. recommends that companies prevent uncontrolled AI sprawl by implementing binding guidelines and providing secure AI services for employees. This way, you maintain control and sustainably reduce risks.

Leading German media outlets are also increasingly warning about the risks of AI, particularly uncontrolled "shadow AI". For instance, heise online reports on the significant rise in shadow AI usage in companies – noting that only around a quarter of firms have clear rules for AI use (Förster, 2025). DER SPIEGEL highlights in a recent report how easily confidential information is put at risk in public chatbots like ChatGPT, exposing companies to substantial data protection and security risks (Böhm & Breithut, 2025). The Süddeutsche Zeitung confirms these warnings, describing AI browsers and freely accessible AI assistants as potential "Trojan horses" capable of harvesting sensitive data, and reporting high levels of shadow AI use in government agencies as well as growing data protection threats from uncontrolled AI tools (Rinaldi, 2025; Ross, 2025; Berlin, 2025).

NENNA is the technical solution that enables companies to protect sensitive data when working with AI, providing a data-secure environment for all AI use cases thanks to its own AI interface with built-in data privacy and additional protection for all AI models in the browser. NENNA complements technical measures with organisational ones, such as AI training, establishing the foundation for a holistically secure, independent, and cost-effective use of AI. This allows companies to use AI applications in compliance with GDPR – the binding European legal framework that applies to all organisations in Germany and the EU.

Using AI safely and to its full extent is a competitive advantage. Read our industry-specific use cases for working with AI here:

Sources:

KPMG International & University of Melbourne. (2025). Trust, attitudes and use of AI: A global study 2025. https://assets.kpmg.com/content/dam/kpmgsites/ae/pdf/trust-attitudes-and-use-of-ai-global-report.pdf.coredownload.inline.pdf

Bitkom e.V. (2025). Press release "Employees increasingly use shadow AI". https://www.bitkom.org/Presse/Presseinformation/Beschaeftigte-nutzen-Schatten-KI

Bitkom e.V. (2025). Guide "Generative AI in Corporate Environments". https://www.bitkom.org/Bitkom/Publikationen/Generative-KI-im-Unternehmen

Förster, M. (2025, 21 October). Shadow AI in companies is increasing significantly. heise online. https://www.heise.de/news/Schatten-KI-in-Unternehmen-nimmt-deutlich-zu-10790550.html

Deloitte. (2025). Cyber Security – Cybersecurity meets AI and Generative AI. https://www.deloitte.com/de/de/services/consulting/perspectives/cybersecurity-meets-ai-genai.html

World Economic Forum. (2025). Global Cybersecurity Outlook 2025. World Economic Forum. https://www.weforum.org/publications/global-cybersecurity-outlook-2025

Tagesspiegel. (2025, 12 November). Shadow AI makes Germany vulnerable. https://background.tagesspiegel.de/it-und-cybersicherheit/briefing/schatten-ki-macht-deutschland-angreifbar

Böhm, M., & Breithut, J. (2025, 26 October). How secure are secrets with OpenAI's chatbot? Der Spiegel. https://www.spiegel.de/netzwelt/web/chatgpt-wie-sicher-sind-geheimnisse-beim-chatbot-von-openai-a-62cdb47d-0ffe-4d97-8ee1-13e2432d128d

Rinaldi, G. (2025, 11 November). Survey: High usage of 'shadow AI' in public authorities. SZ Dossier / Süddeutsche Zeitung. https://www.sz-dossier.de/meldungen/umfrage-hohe-nutzung-von-schatten-ki-in-behoerden-a1a949bf

Ross, M. (2025, 31 October). Why AI browsers turn out to be Trojan horses. SZ Dossier / Süddeutsche Zeitung. https://www.sz-dossier.de/gastbeitraege/warum-sich-ki-browser-als-trojanische-pferde-erweisen-8e058ef2

Berlin, S. (2025, 28 September). The flood of AI assistants is annoying. And dangerous. Süddeutsche Zeitung. https://www.sueddeutsche.de/wirtschaft/kuenstliche-intelligenz-datenschutz-widerspruch-workslop-ki-etikette-li.3316313