MMOOZZGG.com

Neural network API

LLM API security: who sees your requests, from provider and middleman to court

7 min read12 sources

LLM API security depends on who along the path stores your requests. OpenAI keeps API logs for up to 30 days, Anthropic deletes API data within 30 days, and the free tier of the Gemini API gives requests to Google to improve its products and to human reviewers. Below, link by link: retention periods, sources as of 16 October 2026 and a checklist of 7 questions for a supplier.

In short: five places where requests stay

A request to a model passes several links, and a copy may stay at each one. Five places known from documents and public cases:

  • The model provider: logs for abuse monitoring, usually up to 30 days.
  • The free tier: requests go to product improvement, and human reviewers may read them.
  • A middleman or aggregator: its own server and its own policy on top of the provider’s policy.
  • A leak: an open log database, as with DeepSeek in January 2025.
  • A court: an order to keep logs that would otherwise be deleted, as in the New York Times v. OpenAI case.

The provider: retention periods

OpenAI (Your data page, 16 October 2026): API data is not used for training by default. Logs for abuse monitoring are kept up to 30 days, longer if the law requires it. Zero data retention is available only with OpenAI’s approval and not for all API methods.

Anthropic (Privacy Center, articles updated 1 July 2026 and 18 August 2026): API inputs and outputs are deleted within 30 days unless there is a separate zero retention agreement. If a request is flagged as a policy violation, the data is kept up to 2 years. By default, data from commercial products, including the API, does not go to training.

In its privacy policy (updated 10 February 2026), DeepSeek writes that it collects, processes and stores personal data in the People’s Republic of China. The retention period is described in general words: as long as needed to run the service.

The free tier: training and human review

Gemini API Additional Terms (16 October 2026): on the free tier, Google uses requests and answers to improve its products, and human reviewers may read them. Google directly asks you not to send confidential or personal data there.

On the paid tier, requests are not used to improve products, but they are stored for a limited time. The difference between tiers is written in the terms, not in the advertising, so read the terms of the tier your employees actually use.

Employees often start with a free key for a trial. If a client’s contract ends up in such a request, the text is already with the provider under the free tier terms.

Middlemen and aggregators

An aggregator receives your request and passes it to one of the providers. OpenRouter (Provider Logging, 16 October 2026) writes that each provider has its own retention and training policy, and that blocking providers that train models in your settings does not change OpenRouter’s own policy for your requests.

According to its Data Collection page (16 October 2026), OpenRouter does not store requests and answers by default, but it keeps the metadata of each request: the number of tokens, latency. So the chain has at least two policies, and you need to check both. More in our review of API aggregators.

Leaks and courts

On 29 January 2025, Wiz researchers found a DeepSeek database open without a password: more than a million lines of logs with chat history and secret API keys. After the report, DeepSeek quickly closed access.

On 13 May 2025, the court in the New York Times v. OpenAI case ordered OpenAI to keep ChatGPT output logs that would otherwise have been deleted. On 9 October 2025 the duty was lifted for the future, but the logs already kept remained available to the plaintiffs.

A leak can also come from your side. On 2 May 2023 Samsung banned generative AI for its employees after engineers uploaded internal source code to ChatGPT.

The personal data law

Under 152-FZ, personal data of Russian citizens is collected and stored in databases in Russia (Part 5 of Article 18). Since 1 March 2023, the operator files a notice with Roskomnadzor before a cross-border transfer (Part 3 of Article 12).

A request to a foreign API with a client’s name, phone number or passport is such a transfer. Since 30 May 2025, for a repeat leak a legal entity pays 1 to 3% of its revenue, but not less than $300,000 and not more than $6 million.

How we ourselves work with client data is described on the page about data security.

How to check whom you trust

Trust in a supplier is checked through access to the infrastructure, logs and the contract, not through words on a website. Ask these 7 questions to any API supplier, including us, and ask for the answers in writing.

If the answers do not satisfy you, you can keep the model fully in-house: that is how AI in a closed environment works. Our answers to these questions are in the block “Secure API for large volumes” at the bottom of the page and are fixed in the contract.

  • Where are requests physically processed: country, city, data center?
  • How long are requests, answers and metadata stored, and in which document is this written?
  • Are requests used to train models, including on the free tier and in user feedback?
  • Who stands between you and the model, and what is the retention policy of each link?
  • Can you open the logs and the infrastructure yourself, not just get the supplier’s report?
  • Who is the operator of the personal data, is the supplier in the Roskomnadzor register, and will there be a cross-border transfer?
  • What happens to the data in a leak, a court request or the end of the contract: who notifies you, and how soon?

Questions and answers

Do providers use my requests for training? It depends on the tier. As of 16 October 2026, OpenAI and Anthropic write that API data does not go to training by default.

On the free tier of the Gemini API, Google uses requests to improve its products, and human reviewers may read them.

Can you send clients’ personal data to an API? 152-FZ, Russia’s personal data law, requires data of Russian citizens to be collected and stored in databases in Russia, and Roskomnadzor, the data protection regulator, to be notified before a cross-border transfer. A request with a client’s name or phone number to a foreign API is such a transfer, so first check where the supplier processes requests.

Google directly asks you not to send personal data to the free tier of the Gemini API. A repeat leak brings a fine of up to 3% of revenue, but not more than $6 million.

How can you make sure nobody stores your requests? Ask for retention periods in writing and for access to the logs. At OpenAI, logs live up to 30 days, and zero retention is given only on approval.

The New York Times case in 2025 showed that a court can order data that would be deleted under the rules to be kept.

Sources and dates

Contact us