Why run AI in your own account.
Sending company data to a hosted AI service puts it in that service's systems, under terms that can change. These are documented cases: lawsuits, rulings, regulator actions and the companies' own statements. Open cases are described as allegations. Status as of 2026-09-18.
Company data exposed through AI tools.
Some were bugs or breaches at the provider, some were sharing features that made conversations searchable, and one was employees pasting source code into a chatbot.
Samsung source code in ChatGPT
In 2023 Samsung restricted staff use of ChatGPT and similar tools after engineers entered confidential source code into the chatbot.South Korea · May 2023TechCrunch, May 2023
ChatGPT conversation titles and payment details
In March 2023 a software bug let some ChatGPT users see the titles of other users' conversations. It may also have exposed partial payment details for about 1.2 percent of paying subscribers active at the time.Mar 2023 · fixedThe Register, Mar 2023
OpenAI internal forum
The New York Times reported in 2024 that a hacker had accessed an internal OpenAI discussion forum in 2023 and obtained details about its AI technology. OpenAI did not announce the breach publicly at the time.2023 · reported Jul 2024Reuters, Jul 2024
Microsoft AI research storage link
In 2023 security researchers found that a misconfigured link shared by Microsoft AI researchers had exposed 38 TB of internal data, including passwords, keys and internal messages. Microsoft said no customer data was exposed.Disclosed Sep 2023 · fixedWiz Research
DeepSeek open database
In January 2025 researchers found a publicly accessible DeepSeek database containing user chat histories and secret keys. DeepSeek secured it after being notified.Jan 2025 · securedWiz Research
EchoLeak in Microsoft 365 Copilot
In 2025 researchers showed that a single malicious email could have tricked Microsoft 365 Copilot into sending internal company data to an attacker without anyone clicking anything. Microsoft rated the flaw critical and fixed it, and there is no evidence it was exploited.Jun 2025 · patchedCVE-2025-32711Researchers' paper
Shared ChatGPT conversations in search results
In 2025 thousands of shared ChatGPT conversations turned up in Google search results after users ticked a 'discoverable' option. OpenAI removed the option, saying it created too many chances to share things by accident.Jul 2025 · feature removedEngadget, Aug 2025
Shared Grok conversations in search results
In 2025 reporters found that hundreds of thousands of conversations with xAI's Grok chatbot had become searchable on Google through its share links.Aug 2025TechCrunch, Aug 2025
OpenAI customer details through an analytics vendor
In November 2025 OpenAI said an analytics vendor it used had been breached, exposing names and email addresses of some API customers. No chat content or keys were exposed.Nov 2025BleepingComputer, Nov 2025
Deleted ChatGPT conversations kept for a lawsuit
In 2025 a US court ordered OpenAI to keep ChatGPT conversations that users had deleted, as evidence in a copyright lawsuit. The order ended in September 2025, but logs already kept were held for the case.United States · May 2025 · ended Sep 2025Court order, May 2025Engadget, Oct 2025
In Exalt Reserve, the default is open models running in your own cloud account.
Whether any use case also calls an outside AI service is your decision.
Terms that changed, and regulators who stepped in.
Terms of service decide what a provider may do with what you send it, and they can change after you sign up.
Zoom terms of service
In 2023 Zoom faced backlash over terms that appeared to allow it to use customer content to train AI. It rewrote them to state that it does not train AI on customers' calls, chats or files.Aug 2023 · terms revisedZoomThe Record
Slack privacy principles
In 2024 customers found that Slack used workspace data by default to train some of its machine-learning features, and that opting out meant emailing Slack. Slack says it does not train generative AI on customer data.May 2024SlackEngadget, May 2024
Adobe terms of use
In 2024 an update to Adobe's terms led customers to fear their work could be accessed and used for AI. Adobe said it does not train generative AI on customer content and rewrote the terms.Jun 2024 · terms revisedAdobe, Jun 2024
LinkedIn training on member data
In 2024 LinkedIn began using members' data to train generative AI by default, and updated its terms only after this was reported. In November 2025 it extended this to members in the EU, Switzerland, Canada and Hong Kong, again on by default. Opting out does not reverse training already done.Sep 2024 · expanded Nov 2025LinkedInTechCrunch, Sep 2024
Italy's privacy regulator and OpenAI
In 2024 Italy's privacy regulator fined OpenAI 15 million euros, finding it had trained ChatGPT on personal data without an adequate legal basis. A Rome court annulled the fine in 2026 on jurisdictional grounds, without ruling on those findings.Italy · fined Dec 2024 · annulled Mar 2026Euronews, Dec 2024Reuters, Mar 2026
Italy's privacy regulator and DeepSeek
In January 2025 Italy's privacy regulator ordered DeepSeek to stop processing Italian users' data after the company gave what the regulator called insufficient answers about how it handles personal data.Italy · Jan 2025 · order in forceGarante per la protezione dei dati personaliReuters, Jan 2025
FTC on quietly changed terms
The US Federal Trade Commission has warned that quietly changing terms of service to start using customer data for AI training can be an unfair or deceptive practice.United States · Feb 2024Federal Trade Commission
Anthropic consumer terms
In 2025 Anthropic updated its consumer terms so that Claude chats can be used for model training and kept for up to five years unless users turn the setting off. Business and API plans were excluded.Aug 2025Anthropic
WeTransfer terms of service
In 2025 WeTransfer revised its terms after users objected to a clause that mentioned using uploaded files to improve machine-learning models. The company said it does not use AI on shared files.Jul 2025 · terms revisedThe Register, Jul 2025
What you keep
- The cloud subscription and the bill stay yours.
- Data, models, vector stores, and credentials stay in your environment.
- Network, identity, and access policy stay under your control.
- Data is encrypted at rest and in transit.
- Your audit log, kept in your account, with retention you set.
Copyrighted work used to train AI.
Authors, artists, news publishers and record labels have gone to court over AI trained on their work. Courts have ruled both ways, and most cases are still open.
Authors v. Meta (Kadrey)
Authors sued Meta, alleging their books were used without permission to train its Llama models. A US judge ruled for Meta in 2025, and wrote that the ruling did not mean Meta's use was lawful, only that these plaintiffs made the wrong arguments.United States · ruled for Meta Jun 2025 · not yet finalTechCrunch, Jun 2025
Start with an audit
Bring the questions your legal and security teams ask about AI vendors. We go through where your data would go before anything is deployed.