Nearly a third of employees at Russian companies have sent neural networks their logins, passwords, and keys to work folders over the past six months. At the same time, it turns out that ChatGPT doesn't just process requests but also accumulates users' personal data, building a digital dossier on each person. Together, these two factors create a new, underestimated dimension of corporate risk.
What exactly goes into neural networks
An anonymous survey of 1,300 employees at Russian companies using AI in their work, along with nearly 500 executives and specialists from security, IT, HR, and legal fields, produced telling data. Passwords and access credentials turned out to be the most frequent category of confidential information ending up in chatbots — but far from the only one. Over six months, the following were uploaded to neural networks:
- commercial proposals — 25% of respondents;
- financial calculations and reports — 22%;
- work correspondence — 21%;
- contracts — 16%;
- client databases, including those with personal data — 12%.
In the vast majority of cases, the motives are not related to malicious intent. About 20% of respondents were trying to complete a routine task faster. The rest asked the neural network to make sense of a large volume of information or to prepare or edit a document. At the same time, one in five did not understand what work data is even permissible to share with AI services, and some employees did not even realize they were uploading something confidential.
The institutional gap
The key problem is the absence of formalized rules. Neural networks have already become part of everyday work processes, yet employees lack clear regulations defining what can and cannot be sent to chatbots. Companies need not only to write such policies but also to make them convenient to apply — otherwise they will remain on paper.
Particular attention should be paid to the fact that a document sent to a neural network does not disappear after the chat is closed. The service may remember it — and this is no longer a hypothetical risk.
ChatGPT and users' digital portraits
Dedicated cybercrime units directly warn that ChatGPT remembers what it is told and gradually builds a digital dossier on a person. Users share personal experiences with the bot, upload work documents, ask about money and health — and some of this information is stored in the model's memory so it can "recognize" the interlocutor in future sessions.
Academic research documents the scale of the phenomenon: roughly 95% of memory entries are formed without a person's explicit permission, and 91% of accounts have information stored in memory that qualifies as personal data. Digital portraits are built from the accumulated array.
Specialists identify two threats. The first is that a person can be de-anonymized based on the accumulated data, even if they communicate under a pseudonym. The second is that a neural network's memory can be manipulated from outside: it is enough to embed false instructions in a document that the user themselves sends to the bot for processing.
Basic protective measures
- check memory and personalization settings in AI services;
- do not upload work documents or other closed information to the chat;
- remove real names, addresses, and other identifying details from queries;
- discuss personal matters only in temporary chats;
- regularly check what the neural network has remembered.
Calculations, contracts, and client databases sent to a public neural network can no longer be considered internal documents. Their subsequent fate depends on the service's settings, which only a minority of users ever look at.
Cryptalist analysis: The case demonstrates a systemic gap between the speed of generative AI adoption and the maturity of corporate security policies — data leakage occurs not through hacking but through good-faith use of tools. For the market, this is a signal of growing demand for DLP-class solutions and private corporate AI environments, as well as the likely tightening of regulatory requirements for personal data processing in chatbots. Companies that postpone formalizing rules for working with neural networks should view this as a direct warning about reputational and legal costs.