MMOOZZGG.com

AI for lawyers

Data leak through ChatGPT: Samsung, courts, regulators

5 min read6 sources

A data leak through ChatGPT usually starts not with a hack but with an employee pasting an internal document into the chat. That is how it went at Samsung: after engineers uploaded source code to ChatGPT, the company banned generative AI for staff on 2 May 2023. Below are three cases that changed attitudes to public services, the Russian fines for a leak and six protection steps for a law firm.

On this page · 5 sections

Samsung, 2023

Samsung engineers uploaded internal source code to ChatGPT. On 2 May 2023, the company banned employees from using generative AI.

There was no hack: employees themselves sent the code to an outside service to solve a work task faster. For a law firm, the parallel is clear: a client’s contract, a draft claim or a case memo pasted into a public chat to save an hour.

A ban without a replacement does not remove the cause. The task remains, and the employee moves to a personal account or phone, where the firm no longer sees anything.

For a lawyer, the risk is higher than for an engineer: source code has no client names, but a contract does. On top of the client’s trade secret come attorney-client privilege and 152-FZ, Russia’s personal data law.

The NYT lawsuit: keeping chats

In The New York Times case against OpenAI, on 13 May 2025 Judge Ona Wang ordered OpenAI to preserve ChatGPT output logs, including chats that users had deleted.

The duty ended on 26 September 2025. But on 2 December 2025, the court upheld the handover of a sample of 20 million anonymized logs.

The lesson for a lawyer: a deleted chat does not mean deleted data. Users’ conversations can become material in someone else’s court case that the users know nothing about.

In this sense, leaks through ChatGPT go beyond an employee’s mistake: a text can travel further by a court order in another country.

So in the terms of a service, check not only the retention period but also whether the provider can be ordered by a court to keep or hand over conversations.

The fine in Italy

On 20 December 2024, the Italian data protection authority fined OpenAI EUR 15 million for data processing in ChatGPT training.

The provider got the fine, but the question is the same for any firm: on what legal basis is people’s data processed, and do they know about it?

In Russia, the firm answers for its clients’ data as the data operator. Under Part 5 of Article 6 of 152-FZ, if processing is entrusted to another party, the operator, not the service provider, answers to the data subject for that party’s actions.

For the provider, the dispute was about training the model; for a firm, the question is what lawyers send to the service. The firm controls the second question itself.

Fines in Russia for a leak

Since 30 May 2025, under Federal Law 420-FZ, a company is fined for a personal data leak under Parts 12–14 of Article 13.11 of the Code of Administrative Offenses, depending on the number of data subjects:

The ground for a fine under Parts 12–14 is an action or inaction of the operator that led to an unlawful transfer of data. No hack is needed: a document that an employee pasted into an outside service is enough.

1. From 1,000 to 10,000 data subjects: $36,000–$60,000.

2. From 10,000 to 100,000 data subjects: $60,000–$120,000.

3. More than 100,000 data subjects: $120,000–$180,000.

4. A repeated leak: 1–3% of revenue, but no less than $240,000 and no more than $6 million (Part 15); with special categories or biometrics, no less than $300,000 (Part 18).

Separately, Part 11 punishes an operator that did not notify Roskomnadzor, Russia’s data protection regulator, about a leak: $12,000–$36,000. A firm needs to learn in time about every case where a document with personal data went to an outside service without a legal basis.

The number of data subjects in a law firm’s documents grows fast: a register of creditors’ claims or a list of workers in a labor dispute can easily hold thousands of records.

The first commercial court cases on these fines are reviewed in the article on personal data fines.

Six protection steps

Protection against leaks rests on how the work is organized, not on a ban. Six steps for a law firm:

1. AI use policy: which services are allowed, what must not be sent to them and who checks the result.

2. An approved tool instead of personal accounts: a model on the firm’s server or in a cloud in Russia.

3. Anonymizing documents before any transfer outside the firm.

4. Client consent to the use of AI in the legal services agreement.

5. Role-based access: a lawyer sees their own cases, the AI has its own account with rights only to its process, and actions are logged.

6. A check of which services lawyers already use and which documents have gone there.

The technical rules by which we set up AI are in the “How we protect data” block below: we process personal data on servers in Russia under 152-FZ, and text does not go abroad or to services without a data processing agreement.

The firm puts steps 1, 3 and 4 in place itself: these are documents and rules. Steps 2, 5 and 6 need the tool to be set up and a check of what is already happening among the lawyers.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Bloomberg, 02.05.2023 (bloomberg.com)
  2. iTechGuides (review of court documents), December 2025 (itechguides.com)
  3. Euronews, 20.12.2024 (euronews.com)
  4. ConsultantPlus, Federal Law No. 152-FZ of 27 July 2006 “On Personal Data”, Article 6, 27 July 2006 (consultant.ru)
  5. ConsultantPlus, Russian Code of Administrative Offenses, Article 13.11, 4 August 2026 (consultant.ru)
  6. Data security when you roll out AI

Found a mistake? Tell us

Publications

How we protect data

  • Data in Russia

    We process personal data on servers in Russia under 152-FZ, Russia's personal data law. Message text does not go abroad or to services we have no contract with for processing on your behalf.

  • Rights only for its own process

    The agent has its own account in the CRM, the accounting system (1C) and the phone system. A sales agent only needs to create deals; it does not need to export the database or have admin rights.

  • Actions are visible

    Everything the agent does in your CRM goes into the CRM's log. Every change of rules is recorded in the support report, and you can roll it back in a minute.

  • Commands from messages are not carried out

    The agent's instructions forbid changing the rules at the request of an incoming message. If a client asks it to "forget the rules" or to send a list of clients, the request goes to a person.

  • People make the decisions

    The agent does not sign documents, make payments or negotiate. It hands complex and disputed cases to an employee together with the history.

  • The model can be replaced

    We choose the model for a process with regard to where it is deployed, and we can replace it without reworking the rules and connections.

After the rollout, the system, a perpetual license, the data and the access rights are yours. The agent's accounts are created in your systems, and your administrator can revoke any access in a minute.

Data security: all the rules
Contact us