Within Shadow AI
What Happens When Staff Paste Secrets Into AI?
Copying internal text into public AI tools can expose customer, financial, legal and proprietary information outside company controls.
On this page
- Why prompt sharing creates exposure
- Types of business data most at risk
- Safer patterns for summaries and analysis
Page outline Jump by section
Introduction
Many workplace AI risks begin with a simple action: an employee copies internal information into a public AI chatbot to obtain a summary, rewrite, translation, analysis or coding suggestion. The convenience is obvious, but the act can move sensitive information from a controlled corporate environment into systems the organisation does not manage. When approved internal AI tools lag behind employee needs, workers often turn to consumer services, creating a direct pathway for confidential data to leave established security, legal and compliance controls. This is one of the most common and consequential forms of shadow AI because it usually arises from routine productivity tasks rather than deliberate misconduct. Evidence from security researchers, regulators and real-world incidents shows that customer data, source code, financial information and legal documents can all be exposed through seemingly ordinary prompts. [WitnessAI]witness.aiAI5 Enterprise AI Chatbot Security Risks & How to Manage ThemAI5 Enterprise AI Chatbot Security Risks & How to Manage Them - WitnessAIMarch 30, 2026…
What Happens When Staff Paste Secrets Into AI?
The key risk is not that employees intend to disclose information. Rather, they often underestimate what is being shared and where it goes.
A prompt is rarely just a question. To obtain a useful answer, users frequently include background documents, customer correspondence, spreadsheets, meeting notes, software code or contract language. Once submitted, that information may be processed by external providers, retained in logs, transferred across jurisdictions, or handled according to terms that differ from the organisation’s own governance requirements. The result is a loss of visibility and control over sensitive information. [GOV.UK]GOV.UKWithdrawn] Generative AI Framework for HMG (HTML18, 2024…
This exposure can occur even when the employee’s goal is entirely legitimate. Common examples include:
- Pasting a customer complaint into a chatbot to draft a response.
- Uploading source code to diagnose a programming problem.
- Sharing meeting transcripts to create summaries.
- Requesting contract analysis by copying legal text into a public tool.
- Asking an AI assistant to interpret financial reports or forecasts.
Each example transfers information beyond the systems that were originally designed to protect it. [WitnessAI]witness.aiAI5 Enterprise AI Chatbot Security Risks & How to Manage ThemAI5 Enterprise AI Chatbot Security Risks & How to Manage Them - WitnessAIMarch 30, 2026…
A common misunderstanding is that the danger only exists if confidential information later appears publicly. In practice, the risk begins much earlier. Organisations may be unable to verify who processed the data, where it was stored, how long it was retained, whether it crossed national boundaries, or whether it was included in future model improvement workflows. [GOV.UK]GOV.UKWithdrawn] Generative AI Framework for HMG (HTML18, 2024…
Why Prompt Sharing Creates Exposure
The mechanism behind prompt-based data leakage is straightforward but often overlooked.
When an employee submits a prompt, they are not merely consulting a tool installed on their own computer. They are transmitting information to a remote service that may operate under different security, retention and governance arrangements than those approved by their employer. Public AI services vary significantly in how they handle user content, making assumptions especially risky. [OpenAI Help Center]help.openai.comOpen source on openai.com.
Three characteristics make prompt sharing particularly problematic:
The transfer is frictionless. Copying and pasting text takes seconds. Unlike sending files through formal channels, AI interactions often feel conversational and informal, reducing the user’s perception of risk. [WitnessAI]witness.aiAI5 Enterprise AI Chatbot Security Risks & How to Manage ThemAI5 Enterprise AI Chatbot Security Risks & How to Manage Them - WitnessAIMarch 30, 2026…
The amount of information grows quickly. To improve output quality, users often provide additional context. A simple request can evolve into pages of internal material, increasing exposure. [SANS Institute]sans.orgSANS InstituteYour Sensitive Data Has Left the Chat: LLMs as Sensitive Data Detectors | SANS Institute…
The boundaries are unclear. Many employees do not fully understand how different AI services retain, process or use submitted information. Security teams frequently cite this uncertainty as a major governance challenge. [WitnessAI]witness.aiAI5 Enterprise AI Chatbot Security Risks & How to Manage ThemAI5 Enterprise AI Chatbot Security Risks & How to Manage Them - WitnessAIMarch 30, 2026…
The result is a mismatch between the apparent simplicity of the action and its potential organisational consequences.
A Real Example: The Samsung Incident
One of the most widely cited cases occurred in 2023 when Samsung employees reportedly entered confidential company information into ChatGPT while seeking assistance with work tasks. Information disclosed through prompts included source code and internal meeting content. The incident became a prominent example of how routine AI use can create confidentiality risks when sensitive information is submitted to external systems. [OECD.AI]oecd.aiSamsung Employees Leak Sensitive Data to ChatGPT, Raising AI Confidentiality Concerns - OECD.AIApril 6, 2023…
The significance of the case was not merely the specific data involved. It demonstrated how quickly an ordinary productivity workflow can become a corporate security issue. Employees were attempting to solve problems more efficiently, yet the process moved proprietary information outside approved organisational controls. [OECD.AI]oecd.aiSamsung Employees Leak Sensitive Data to ChatGPT, Raising AI Confidentiality Concerns - OECD.AIApril 6, 2023…
The incident also highlighted an important lesson: data leakage through AI prompts is often accidental. The most serious exposures frequently arise from convenience rather than malicious intent. [OECD.AI]oecd.aiSamsung Employees Leak Sensitive Data to ChatGPT, Raising AI Confidentiality Concerns - OECD.AIApril 6, 2023…
Types of Business Data Most at Risk
Not all information carries equal risk when entered into public AI tools. Several categories appear repeatedly in security guidance and incident analyses.
Customer and Personal Information
Customer records, contact details, support tickets, employee records and other personally identifiable information can create privacy and regulatory concerns. In many jurisdictions, organisations have legal obligations governing how such information is processed and transferred. [OWASP]owasp.orgOWASP Top 10 for…
Financial Information
Budgets, forecasts, pricing models, acquisition plans and earnings-related information can reveal commercially sensitive details about a business. Exposure may create competitive, legal or reputational risks. [OWASP]owasp.orgOWASP Top 10 for…
Proprietary Code and Technical Designs
Developers often use AI tools for debugging and code review. However, source code, architecture diagrams and engineering specifications may represent valuable intellectual property. The Samsung case illustrated how quickly such material can enter external systems. [OECD.AI]oecd.aiSamsung Employees Leak Sensitive Data to ChatGPT, Raising AI Confidentiality Concerns - OECD.AIApril 6, 2023…
Legal and Contractual Documents
Contracts, litigation materials, regulatory filings and privileged communications may contain sensitive obligations, strategies or legal advice. Sharing such material with unauthorised systems can create governance and confidentiality concerns. [OWASP]owasp.orgOWASP Top 10 for…
Credentials and Security Information
Passwords, access tokens, internal system details and security configurations are particularly dangerous because exposure can create immediate operational risks. Security frameworks increasingly classify sensitive information disclosure as a major large-language-model vulnerability category. [OWASP]owasp.orgOWASP Top 10 for…
Safer Patterns for Summaries and Analysis
The challenge is not that employees need summaries, translations or analytical assistance. The challenge is obtaining those benefits without exposing unnecessary information.
A safer approach is to minimise the amount of sensitive content included in prompts. Instead of sharing complete documents, users can often remove identifying details, replace names with placeholders, redact account numbers and provide only the specific sections relevant to the task. This reduces the volume of information leaving the organisation while still allowing useful assistance. [OWASP]owasp.orgOWASP Top 10 for…
Another effective pattern is to separate the analytical question from the underlying data. Rather than uploading a confidential contract and asking for recommendations, a user can describe the contractual issue in abstract terms and request a framework for analysis. The AI provides guidance without receiving the sensitive document itself.
Many organisations also deploy enterprise AI platforms with contractual privacy protections, logging controls and restrictions on model training. Business-oriented AI services often operate under different data-handling arrangements than consumer offerings, making tool selection a governance issue rather than merely a technical one. [OpenAI Help Center]help.openai.comOpen source on openai.com.
Finally, employees benefit from clear decision rules. A simple question can prevent many incidents: Would I be comfortable sending this exact information to an external third party under company policy? If the answer is uncertain, the material probably should not be pasted into a public AI chatbot.
Why This Risk Persists
Despite growing awareness, prompt-based data leakage remains difficult to eliminate because it exploits normal workplace behaviour. Employees are rewarded for speed, efficiency and problem-solving. Public AI tools often provide immediate value, while approved alternatives may be unavailable or less capable.
This creates a recurring tension at the heart of shadow AI. Workers see a useful assistant; security and compliance teams see an uncontrolled data-transfer channel. Neither perspective is entirely wrong. The risk emerges when convenience obscures the fact that a prompt is also a disclosure event.
As organisations adopt AI more broadly, understanding that distinction becomes essential. The most damaging data leak may not begin with a sophisticated cyberattack. It may begin with a helpful request copied into the wrong chatbot. [WitnessAI+2SANS Institute]witness.aiAI5 Enterprise AI Chatbot Security Risks & How to Manage ThemAI5 Enterprise AI Chatbot Security Risks & How to Manage Them - WitnessAIMarch 30, 2026…
Amazon book picks
Further Reading
Books and field guides related to What Happens When Staff Paste Secrets Into AI?. Use these as the next step if you want deeper reading beyond the article.
Principles of AI Governance and Model Risk Management
Addresses controls around AI use, risk management, and responsible deployment.
Transformational Security Awareness
Explains human behavior risks that lead to data exposure.
The Phoenix Project
Illustrates governance, process discipline, and consequences of uncontrolled information flows.
Cybersecurity and Cyberwar
Provides accessible context on information security and organizational risk.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: owasp.org
Link: https://owasp.org/www-project-top-10-for-large-language-model-applications/assets/PDF/OWASP-Top-10-for-LLMs-v2025.pdfSource snippet
OWASP Top 10 for...
-
Source: GOV.UK
Title: [Withdrawn] [Generative AI]({{ ‘generative-ai/’ | relative_url }}) Framework for HMG (HTML)
Link: https://www.gov.uk/government/publications/generative-ai-framework-for-hmg/generative-ai-framework-for-hmg-html/Source snippet
18, 2024...
-
Source: sans.org
Link: https://www.sans.org/white-papers/your-sensitive-data-has-left-llms-sensitive-data-detectorsSource snippet
SANS InstituteYour Sensitive Data Has Left the Chat: LLMs as Sensitive Data Detectors | SANS Institute...
-
Source: help.openai.com
Link: https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance%23.pict -
Source: oecd.ai
Link: https://oecd.ai/en/incidents/2023-04-06-93c9Source snippet
Samsung Employees Leak Sensitive Data to ChatGPT, Raising AI Confidentiality Concerns - OECD.AIApril 6, 2023...
Published: April 6, 2023
-
Source: help-lb.openai.com
Title: What is Chat GPT? | Open AI Help Center
Link: https://help-lb.openai.com/en/articles/6783457-what-is-chatgpt -
Source: witness.ai
Title: AI5 Enterprise AI Chatbot Security Risks & How to Manage Them
Link: https://witness.ai/blog/chatbot-security-risks/Source snippet
AI5 Enterprise AI Chatbot Security Risks & How to Manage Them - WitnessAIMarch 30, 2026...
Published: March 30, 2026
Additional References
-
Source: reddit.com
Link: https://www.reddit.com/r/ChatGPT/comments/17mugdeSource snippet
If ChatGPT is stateless and does not update its underlying model based on user prompts, how did the Samsung data breach occur?...
-
Source: arxiv.org
Title: Mind your key: An Empirical Study of LLM API Credential Leakage in i OS Apps
Link: https://arxiv.org/abs/2606.12212Source snippet
Mind your key: An Empirical Study of LLM API Credential Leakage in iOS AppsJune 10, 2026...
Published: June 10, 2026
-
Source: ftc.gov
Title: www.ftc.gov Privacy and Security Enforcement | Federal Trade Commission
Link: https://www.ftc.gov/news-events/topics/protecting-consumer-privacy-security/privacy-security-enforcementSource snippet
and Security Enforcement | Federal Trade Commission...
-
Source: youtube.com
Title: www.youtube.com OWAS P LLM Security 02: Sensitive Information Disclosure
Link: https://www.youtube.com/watch?v=QyerdsajWT0Source snippet
LLM Security 02: Sensitive Information Disclosure - YouTubeJanuary 18, 2026...
Published: January 18, 2026
-
Source: youtube.com
Title: Data Security in the Age of AI
Link: https://www.youtube.com/watch?v=BZSRgt-ScaISource snippet
Five AI Risks That Can Get You Fired—And How to Avoid Them...
-
Source: youtube.com
Title: Acronis Gen AI Protection
Link: https://www.youtube.com/watch?v=JgD2GN1J_h8Source snippet
Combat Shadow AI with CSG | The cyber security experts...
-
Source: youtube.com
Title: Data & AI Security Explained: Protect Sensitive Data in the Age of AI
Link: https://www.youtube.com/watch?v=407uMH-S85USource snippet
Acronis GenAI Protection...
-
Source: youtube.com
Title: Combat Shadow AI with CSG | The cyber security experts
Link: https://www.youtube.com/watch?v=OXtlY02IDlMSource snippet
Data Security in the Age of AI...
-
Source: trusttrace.io
Title: Trust Trace OWASP LLM Top 10 — Trust Trace Docs
Link: https://www.trusttrace.io/docs/owasp -
Source: youtube.com
Title: Five AI Risks That Can Get You Fired—And How to Avoid Them
Link: https://www.youtube.com/watch?v=1m55T8xST9s
Topic Tree