What is AI abuse monitoring? What seven AI providers keep, and for how long

The Morse teamPublished November 10, 202610 min read

AI abuse monitoring is the process an AI provider uses to scan the prompts and replies sent to its models for uses that break its rules, keeping a copy of some or all of them so its systems, and sometimes its staff, can check what was flagged. As of November 2026, OpenAI, xAI, and Mistral keep API prompts and replies for up to 30 days for it, Google keeps Gemini API prompts for 55 days, Anthropic keeps flagged API content for up to 2 years, and Microsoft and Amazon store prompts only when they're flagged or sent to certain models.

We read each provider's own documentation on November 10, 2026. Abuse monitoring is why "we don't train on your data" and "we don't keep your data" are different promises. A provider can keep the first and still hold your prompts for weeks. Most providers say plainly how long they keep these logs, and Microsoft and Google go furthest in saying who reads flagged prompts and how those people get access.

How long do AI providers keep data for abuse monitoring?

Anywhere from nothing to two years, depending on the provider, the model, and your contract. Each cell links to the provider's own page, read November 10, 2026.

Six keep every prompt for abuse checks by default: OpenAI, xAI, and Mistral for up to 30 days, Google's Gemini API for 55, and Google Cloud for up to 30 on its Frontier Models and on "Advanced AI" partner models. Anthropic's API deletes inputs and outputs within 30 days as its standard retention, and keeps a longer copy only of what its systems flag. Microsoft, and Google Cloud on Google's standard models, store a prompt only after a classifier flags it. Anthropic stands apart on duration, since a flagged prompt can outlive the 30-day window by almost two years.

What is OpenAI abuse monitoring?

OpenAI abuse monitoring is a log of API traffic that OpenAI keeps for up to 30 days to look for misuse. OpenAI's data controls page says the logs are "generated for all API feature usage" and "may contain certain customer content, such as prompts and responses, as well as metadata derived from that customer content, such as classifier outputs." OpenAI keeps them longer if "required by law, or is reasonably necessary to protect our services or any third party from harm."

OpenAI offers two ways out, both by approval through its sales team. Zero Data Retention keeps customer content out of the logs. Modified Abuse Monitoring "excludes customer content (other than image and file inputs in rare cases)" from the logs "while still allowing the customer to take advantage of the full capabilities of the OpenAI platform." Neither covers everything. If OpenAI's classifier detects potential child sexual abuse material in an image, "the image will be retained for manual review," whichever control is on. OpenAI also describes Private Safety Processing, which lets it "perform automated safety monitoring while preserving Zero Data Retention protections." We cover OpenAI's other API retention periods in is the OpenAI API private.

What is Azure OpenAI abuse monitoring?

Azure OpenAI abuse monitoring is Microsoft's own system for the OpenAI models it hosts, and it runs separately from OpenAI's. Microsoft's data privacy page says prompts and completions sent to Azure are "NOT available to OpenAI." Microsoft's system works in steps, according to its abuse monitoring page:

1. Classifiers score each prompt and completion for harmful content.
2. Pattern detection scores an account's usage, weighing how often and how severely harmful content appears and signs of intent, "such as recurring jailbreak attempts."
3. Review of flagged items is done "by automated means including AI models such as LLMs" by default. Prompts in automated review "are not stored by the abuse monitoring system."
4. Human review comes in only when automated review falls short of "applicable confidence thresholds" or isn't available.

The human-review copies sit in a data store kept in the Azure geography where your resource is deployed, separated by resource. Reviewers are "authorized Microsoft employees" who reach the data through Secure Access Workstations, with "Just-In-Time (JIT) request approval granted by team managers," and for deployments in the European Economic Area, the reviewers are located there too.

Microsoft doesn't say how long it keeps them: none of Microsoft's three pages on the subject, read November 10, 2026, gives a number of days. The data privacy page was last updated June 5, 2026.

What is modified abuse monitoring?

Modified abuse monitoring is Microsoft's exemption from storage and human review, for customers who can't let Microsoft keep their prompts. Microsoft's limited access page says it's "available only to customers and partners managed by a Microsoft account team or under an eligible program," who apply by form. It adds that "some advanced Models sold by Azure may have more stringent criteria."

Approval doesn't switch monitoring off. Microsoft says automated review "may still be conducted" as prompts arrive, and a customer whose traffic shows "severe or recurring abuse" may be asked to turn human review back on. You can check the setting yourself: on the resource's Overview page, open JSON view and look for a capability named "ContentLogging" set to "false." The attribute appears only when storage for abuse monitoring is off. Microsoft warns that without human review, "detection of potential abuse may be less accurate."

Does Google log Gemini prompts for abuse monitoring?

Yes, on both of its platforms, and the Gemini API keeps prompts and output for 55 days (Google), as we cover in is the OpenAI API private.

Google Cloud, where Vertex AI now goes by Gemini Enterprise Agent Platform, has three tiers on its abuse monitoring page, last updated November 10, 2026:

• Google's standard models log prompts only after automated classifiers find suspicious activity, for "up to 90 days" in your project's region. Customers with a Google Cloud Master Agreement "are exempt from prompt logging for this abuse monitoring by default."
• Google Frontier Models, such as Gemini 4 Carbon, log prompts and responses for up to 30 days "inside the customer's environment." Human review applies only to customers on Google Cloud's online terms.
• "Advanced AI" partner models log every prompt and response for up to 30 days. Google's list includes Claude Mythos and Claude Fable, and OpenAI's GPT Astra, Sol, Terra, and Luna families.

Google notes that abuse-monitoring logs "are not encrypted by Customer-managed encryption keys (CMEK)" in any of the three tiers.

Does the same model get monitored differently on each cloud?

Yes. An OpenAI model is logged four different ways for abuse, depending on where you call it:

Claude's most capable models split the same way. Anthropic itself has kept prompts and outputs on its covered, Mythos-class models for 30 days since June 9, 2026, and says "no Anthropic personnel can read your retained conversations" by default (Anthropic). On Bedrock, Amazon keeps all traffic to Claude Fable 5, 5.1, and 5.2 for up to 30 days, may review flagged traffic itself, and says the data is "not shared with third-party model providers." Google Cloud goes the other way. To use Claude Fable 5 or Mythos 5 there, Google says "you must enable sharing this data with Anthropic for abuse monitoring" (Google Cloud).

Does zero data retention turn off abuse monitoring?

Mostly, with carve-outs every provider writes for itself. Anthropic's zero data retention agreement excludes retention needed to "comply with law or combat misuse or harm," and Anthropic "still retains User Safety classifier results" (Anthropic). Mistral grants it only for stateless calls on pay-as-you-go, after you "provide sufficient detail of your legitimate reasons" (Mistral). For how each provider defines zero data retention, see what is zero data retention. For who reviews chats in the consumer apps, see who reviews AI chats.

How Morse keeps your prompts out of an abuse log under your name

Every abuse log on this page is tied to the customer account that sent the traffic. Microsoft tells that customer by email (Microsoft), Google may pause its access (Google), and AWS writes to the address on the account (AWS). With your own API key, or in an app that sends your messages under your name, that account is you, and the copies a provider keeps carry your name. Morse keeps your name off those copies, in a different way in each of its three privacy modes.

• Anonymous, for GPT, Claude, and Gemini without your name in the provider's abuse log. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account, so the provider answers without knowing it's you. The provider still sees the message and may keep it for a limited time to check for abuse, under its own policy, so its abuse checks still apply. Anonymous mode changes whose account a kept copy is filed under, so leave out names and details in the message itself that would point to you. Our guide to using ChatGPT anonymously covers the rest.
• Private, the default, for chats you'd rather no host keep. Open-weight models like DeepSeek, Kimi, and GLM, plus xAI's Grok, run under the zero data retention agreements Morse has with its inference, search, and other providers, so they keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't. Otherwise Morse relies on the provider's commitment.
• E2EE, for prompts no provider should read at all. Your browser checks a secure enclave's hardware attestation, then encrypts the message so only that verified enclave can open it. Morse's servers relay ciphertext they can't read, and the message is readable only inside that enclave. The enclaves are run by third-party providers, not by Morse, and the guarantee depends on the enclave hardware and attestation working as designed. E2EE runs open-weight models only, without web search or image generation.

In every mode, your chat history is stored only in your browser, encrypted, and Morse's servers don't keep your messages or log what you write. Morse doesn't train AI on your chats, sell your data, or show ads, and because it doesn't store chats, it usually has very little to give authorities (privacy policy). Clearing your browser deletes your chats for good, and Morse can't recover them, so save anything you need.

Try a chat from the Morse home page without an account, or start free with no card needed. E2EE chats come with Pro, at $19 a month.