---
id: CP-306
slug: shadow-ai-public-chatbot-data-leak
lang: en
title: "Shadow AI: pasting company data into a free public chatbot"
description: Pasting a work document into a public chatbot means handing it to a third party. What NCSC, ANSSI and CNIL say, and what to do instead.
canonical_url: https://startbrain.io/cyberpedia/en/articles/shadow-ai-public-chatbot-data-leak/
markdown_url: https://startbrain.io/cyberpedia/en/articles/shadow-ai-public-chatbot-data-leak.md
publisher: Brain (https://startbrain.io)
license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
cite_as: "Brain, \"Shadow AI: pasting company data into a free public chatbot\", Cyberpedia, 2026. https://startbrain.io/cyberpedia/en/articles/shadow-ai-public-chatbot-data-leak/ (CC BY 4.0)"
wing: 4
theme: authorized_tools
reflex: air-019
schema: G5
updated: "2026-11-14"
translations:
  en: https://startbrain.io/cyberpedia/en/articles/shadow-ai-public-chatbot-data-leak/
  fr: https://startbrain.io/cyberpedia/fr/articles/shadow-ia-chatbot-grand-public-fuite-donnees/
  de: https://startbrain.io/cyberpedia/de/articles/shadow-ki-oeffentlicher-chatbot-datenabfluss/
  es: https://startbrain.io/cyberpedia/es/articles/shadow-ia-chatbot-gran-publico-fuga-datos/
  it: https://startbrain.io/cyberpedia/it/articles/shadow-ia-chatbot-grande-pubblico-fuga-dati/
  nl: https://startbrain.io/cyberpedia/nl/articles/shadow-ai-chatbot-groot-publiek-datalek/
  pl: https://startbrain.io/cyberpedia/pl/articles/shadow-ai-ogolnodostepny-chatbot-wyciek-danych/
  pt: https://startbrain.io/cyberpedia/pt/articles/shadow-ia-chatbot-publico-fuga-dados/
  ja: https://startbrain.io/cyberpedia/ja/articles/shadow-ai-public-chatbot-data-leak/
sources:
  - "n": 1
    title: "ChatGPT and large language models: what's the risk?"
    publisher: NCSC (UK)
    url: https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk
    published: "2023-03-14"
    accessed: "2026-11-14"
  - "n": 2
    title: Security recommendations for a generative AI system
    publisher: ANSSI
    url: https://messervices.cyber.gouv.fr/documents-guides/security_recommandations_for_a_generative_ai_system.pdf
    published: "2024-04-29"
    accessed: "2026-11-14"
  - "n": 3
    title: "Meet “AI”, your new colleague: Could it expose your company's secrets?"
    publisher: ESET
    url: https://www.welivesecurity.com/2023/05/17/meet-ai-new-colleague-work-company-data/
    published: "2023-05-17"
    accessed: "2026-11-14"
  - "n": 4
    title: Les questions-réponses de la CNIL sur l’utilisation d’un système d’IA générative
    publisher: CNIL
    url: https://www.cnil.fr/fr/les-questions-reponses-de-la-cnil-sur-lutilisation-dun-systeme-dia-generative
    published: "2024-07-18"
    accessed: "2026-11-14"
  - "n": 5
    title: Comment déployer une IA générative ? La CNIL apporte de premières précisions
    publisher: CNIL
    url: https://www.cnil.fr/fr/comment-deployer-une-ia-generative-la-cnil-apporte-de-premieres-precisions
    published: "2024-07-18"
    accessed: "2026-11-14"
  - "n": 6
    title: LLM02:2025 Sensitive Information Disclosure
    publisher: OWASP
    url: https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/
    published: "2024-04-10"
    accessed: "2026-11-14"
---

# Shadow AI: pasting company data into a free public chatbot

A public chatbot answers fast and asks for nothing in return. Paste in meeting minutes, a piece of code or a customer list, and you get a summary, a fix or a translation. This harmless-looking gesture has an effect the screen does not show: the text leaves the company and lands with a third party.

Shadow AI is when an employee uses an AI tool that their organisation has not approved. The NCSC (UK) notes that the query is visible to the organisation providing the model [1]. ANSSI [2] puts these services among those to which sensitive data must not be sent.

## In 30 seconds

- Where it happens: in a chatbot open to the public, used outside the tools the company has approved, to summarise, translate, correct or write code.
- The risk: what you paste is visible to the provider, which may keep it and use it [1]. The company loses control of the data.
- The reflex: paste nothing confidential into a public chatbot [2].

## Straight answers

### Does the chatbot learn what I type and repeat it to others?

The NCSC (UK) [1] separates two things. At the date of its post (March 2023), a model does not automatically add queries to what it knows, so entering a piece of information does not make it part of the model for other users [1]. The query, however, is visible to the organisation providing the model, it is stored, and it will almost certainly be used at some point to develop the service or the model [1]. The provider, its partners or its contractors could therefore read the queries [1].

ESET [3] adds that OpenAI introduced, at the end of April 2023, the option to turn off conversation history in ChatGPT. According to OpenAI, quoted by ESET, conversations started without history are not used to train or improve its models [3].

### Which data must never leave?

ANSSI [2] states that sensitive data must absolutely not be sent to third-party generative AI services such as ChatGPT, Gemini, Copilot, DeepL (text translation) or Perplexity. It lists restricted-distribution or classified data, research work covered by the protection of the nation's scientific and technical potential, personal data, the company's contractual, legal and financial data, and IT secrets such as passwords or API keys [2].

OWASP [6] gives a similar list of sensitive information: personal data, financial details, health records, confidential business data, security credentials and legal documents [6].

### Why does ANSSI compare this to storage at a vendor?

Because, according to ANSSI [2], sending text, images or documents to a publicly accessible generative AI service amounts to placing them on a storage space owned by the vendor. Isolation between customers and protection of the confidentiality of the data sent are not always state of the art, and rest solely on trust in the provider [2]. With most services, the data sent is collected and used by the provider to optimise its models [2].

### Is there any use of these tools that remains acceptable?

CNIL [4] answers with conditions. For non-confidential uses, a public service can be considered with appropriate safeguards, provided that accounts created with personal email addresses are avoided and, where relevant, the provider's ability to reuse usage data is turned off [4]. As soon as customers' or staff members' personal data, or sensitive or strategic documentation, is involved, CNIL says it generally seems more appropriate and safer to favour an on-site deployment [4].

## How it works

The path of data pasted into a chatbot is simple. The text goes to the organisation that provides the model. That organisation stores it and may use it to develop its service [1]. The employee keeps no control afterwards: for ANSSI [2], the customer company does not control the service, and it is impossible to guarantee that the confidentiality of the data entered meets its security requirements [2].

The NCSC (UK) [1] cites other risks. Queries stored online may be hacked, leaked or, more likely, accidentally made public, sometimes with information that identifies the user [1]. The operator of the model may also be bought later by an organisation whose privacy policy differs from the one in force when the data was entered [1]. Finally, a question's sensitivity may come from the person asking it or from the moment, and putting together several queries made with the same account reveals more than each one alone [1].

OWASP [6] describes the other direction of the risk: data given to a model without a second thought may come back out later in its answers. It asks users to understand this danger [6].

## What you do

- Paste nothing confidential into a public chatbot. CNIL [4] gives the example of personal data and company data, especially data covered by trade secrecy or by professional-conduct obligations [4]. The NCSC (UK) [1] recommends not including sensitive information in queries to public models and not submitting a query that would be a problem if it became public [1].
- Use the tool the company has approved. CNIL [5] recommends framing uses with a list of authorised and prohibited uses, for example not providing personal data to the system [5]. It also recommends that organisations set these uses in internal policies or charters [4]. ESET [3] says the same: the company's policy should define in which circumstances employees may use these tools [3].
- Review what the tool produces before using it. Users should take a critical view of outputs, and a good practice is never to reproduce them as they are [4].
- If access to a tool goes through an account, use the professional address provided rather than a personal one, within the framework the organisation has authorised [4].

## Pitfalls

The first pitfall is convenience. ANSSI [2] notes that these tools are so easy to use that people are tempted to entrust work data to them, for example to translate a text [2].

The second is believing that a setting is enough. For OWASP [6], restrictions added to a model's internal instructions about the types of data to return may reduce the risk, but they are not always honoured and can be bypassed by prompt injection or other methods [6]. The option to turn off history mentioned above concerns the training of models [3]. The NCSC (UK) reminds us, for its part, that the query is visible to the organisation providing the model [1].

The third is thinking the provider guarantees security. ESET [3] recalls that OpenAI had to take ChatGPT offline on 20 March 2023. OpenAI said on 24 March that a bug in an open-source library had let some users see titles from another active user's history, and that the payment information of 1.2% of ChatGPT Plus subscribers active during a nine-hour window may have been made visible by mistake [3].

The fourth is believing that general advice replaces a rule. CNIL [4] specifies that, when the provider may reuse data under its terms of use, the organisation decides case by case whether to ban all data supply or only certain categories [4]. The rule that applies at your desk is your organisation's.

## Real cases

### Samsung, March 2023: code and a meeting recording in ChatGPT, according to ESET

ESET [3] reports that, at the end of March 2023, the South Korean outlet The Economist Korea reported three independent incidents at Samsung Electronics [3]. According to this account, the company had asked its employees to be careful about the information entered in their queries, but some reportedly let internal data slip while using ChatGPT [3].

One employee reportedly entered faulty source code related to the measurement database of semiconductor equipment in order to look for a fix [3]. A second reportedly did the same with a program for identifying defective equipment, to get the code optimised [3]. A third reportedly uploaded the recordings of a meeting to get minutes [3]. ESET states that Samsung plans to develop its own internal AI service to help its employees in their work [3].

OWASP [6] lists among its references an article titled "Lessons learned from ChatGPT’s Samsung leak" [6].

## Sources

1. NCSC (UK), "ChatGPT and large language models: what's the risk?", 14 March 2023. https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk (accessed 14 November 2026)
2. ANSSI, "Security recommendations for a generative AI system", 29 April 2024. https://messervices.cyber.gouv.fr/documents-guides/security_recommandations_for_a_generative_ai_system.pdf (accessed 14 November 2026)
3. ESET, "Meet “AI”, your new colleague: Could it expose your company's secrets?", 17 May 2023. https://www.welivesecurity.com/2023/05/17/meet-ai-new-colleague-work-company-data/ (accessed 14 November 2026)
4. CNIL, "Les questions-réponses de la CNIL sur l’utilisation d’un système d’IA générative", 18 July 2024. https://www.cnil.fr/fr/les-questions-reponses-de-la-cnil-sur-lutilisation-dun-systeme-dia-generative (accessed 14 November 2026)
5. CNIL, "Comment déployer une IA générative ? La CNIL apporte de premières précisions", 18 July 2024. https://www.cnil.fr/fr/comment-deployer-une-ia-generative-la-cnil-apporte-de-premieres-precisions (accessed 14 November 2026)
6. OWASP, "LLM02:2025 Sensitive Information Disclosure", 10 April 2024. https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/ (accessed 14 November 2026)

---

Original source: Cyberpedia, the public cybersecurity and AI encyclopedia published by Brain. Canonical version: https://startbrain.io/cyberpedia/en/articles/shadow-ai-public-chatbot-data-leak/
You may reuse this text freely under the CC BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), provided you credit "Brain, Cyberpedia" with a link to the canonical URL above.
