The Shtab blog published an analysis of the risks associated with sending corporate information to external AI services. The practical takeaway is simple, but not merely decorative: security depends not only on the model you choose, but also on what data you send it, who has access to that data, and what happens after the request.

The author suggests assessing data according to two factors: how much harm a leak would cause and whether its consequences could be remedied. Public company materials add almost no risk, since they are already publicly available. Internal documents without trade secrets or personal data pose a more moderate risk, while source code, product plans, and financial reports could give competitors an advantage.

The author identifies personal data, medical, banking, and state secrets as the most sensitive. For organizations that handle such information, the consequences may include legal liability, financial losses, and reputational damage. In the case of state secrets, the risk also extends to employees with the appropriate clearance. These are not categories for a “let’s paste it into the chat first and figure it out later” kind of experiment.

The risk doesn’t arise only during model training

One mechanism is using user conversations to train models. The article says this option is usually enabled in consumer versions of popular services, while enterprise products may offer the option to turn it off. But an employee might work on a job-related task using a personal account and fail to check the settings. In that case, corporate policy on paper does nothing to change the route by which data leaves the company.

As an example, the author cites the case of Samsung: in March 2023, engineers were allowed to use ChatGPT, and three incidents occurred within the first three weeks. Employees sent the service source code, code for detecting defective equipment, and a meeting transcript. According to the article, this data ended up on OpenAI’s servers and could have been used for training. Samsung later restricted the use of generative AI on work devices and its internal network, but that measure could not undo transfers that had already taken place.

There is another class of problem: provider errors can make chats available to other users or search engines. The article mentions a ChatGPT outage in 2023 during which the titles of other people’s conversations and some subscribers’ personal and payment information were exposed. It also describes instances of ChatGPT conversations being indexed in search in July 2025, followed later by similar situations involving Grok and DeepSeek. The author links these to chat-sharing features: an option that looks like an ordinary link can broaden a conversation’s audience far beyond what the user expected.

Sending data to a vendor makes you dependent on its procedures for storing, accessing, and processing that data. The article also mentions The New York Times’ legal dispute with OpenAI over the use of articles for training. According to the author, during the proceedings the court required conversation logs to be preserved, including those deleted by users, and later ordered a sample of 20 million anonymized conversations to be provided. This example shows that deleting a chat in the interface does not necessarily mean the data has disappeared from all systems or contexts.

Agents add a separate attack surface when they are given access to email, drives, or corporate correspondence. The article describes a vulnerability found in 2025 in Microsoft 365 Copilot: malicious emails containing hidden instructions could prompt an assistant with email access to gather sensitive context and send it to an external server. Microsoft fixed the issue on its end; the article provides no information about whether anyone managed to exploit it. It’s important here not to confuse integration convenience with a free expansion of trust: every additional access permission needs to be justified.

What a company can do

The first option is to deploy a language model within the company’s own environment or on dedicated infrastructure. This approach reduces the need to send requests to a third-party service, but requires resources for equipment and maintenance. For a company without a suitable team, that can be a significant obstacle. An alternative is an enterprise API with clear contractual terms for data storage; the fact that it’s an enterprise plan alone does not explain exactly where requests are processed or what rules the provider applies.

The second measure is to minimize and anonymize input data. Instead of a full report, you can send the model only the relevant excerpt; remove names, project titles, and financial figures if they aren’t needed for the task. This reduces the amount of information that could be exposed, though it does not eliminate the risk entirely. This approach is more complicated for code: even without names or comments, its structure may reveal commercially important information. Less context may also worsen the answer, so the result needs to be checked.

A proxy between the employee and the model can serve as a technical supplement. It can detect common types of sensitive data—for example, card numbers, passport details, or API keys—and mask or block the request. Such a gateway won’t recognize every possible wording and is no substitute for data classification, but it adds a layer of control where human error is entirely predictable. Automation does not remove the need to monitor access policies and review usage logs.

Human review is needed both before and after consulting the model. Before sending a request, assess whether its contents can really be shared; afterward, check the facts and don’t let fabricated claims make their way into an email, document, or publication. The article gives examples of chatbots producing incorrect answers, including the Air Canada case in 2024, when a bot told a passenger about a discount that wasn’t in the company’s rules, and the dispute ended up in court.

A ban on AI by itself does not guarantee that employees will stop using it: a personal phone and account remain outside the work device. So the author advises against relying on bans alone and recommends providing a controlled tool with appropriate access permissions. The irony here is purely engineering: a policy that can be bypassed with one smartphone looks good in a document, but works poorly as protection.

Before choosing a service, companies should define data classes, check which models the product uses, and read the terms for processing, storage, and access to chats. They should also find out whether the assistant inherits the user’s permissions: if it doesn’t, it may show employees materials they shouldn’t be able to see. For critical data, the basic criterion should not be a promise of “safe AI,” but a clear processing chain and a defined amount of information the company is willing to send beyond its own walls.