Zero retention in AI gateways: contract requirements

A zero retention policy in an AI gateway does not guarantee the absence of screening. Subsequent retention depends on the terms of each layer. Before signing, the buyer should request the screening, review, access, and preservation policies of the providers involved in writing.
On November 16, 2026, OpenAI announced an image classifier to identify possible child sexual abuse material, known as CSAM. The announcement does not describe every gateway's policy. It helps formulate a question that a generic zero retention promise does not answer: what happens between the arrival of a request and its eventual deletion?
Decision matrix / Key facts:
- Screening during processing and retaining data afterward are different decisions.
- OpenAI documents an exception for images flagged as possible CSAM.
- A flag is not a conclusion and does not, by itself, constitute a report to CyberTipline.
- The contract needs to identify policies and exceptions by provider and by route.
Does zero retention in an AI gateway prevent screening?
No. Zero retention in an AI gateway can limit subsequent storage without preventing screening. On November 16, 2026, OpenAI announced an image classifier. Its zero retention policy provides for an exception for images flagged as possible CSAM. Each provider has its own terms.
On August 19, 2026, OpenAI said its zero retention policy for eligible API customers, relevant to AI gateways, does not retain prompts or responses after processing, except for images flagged as possible CSAM. The classifier announcement does not date the start of the exception. OpenAI describes Private Safety Processing: content remains on customer-controlled infrastructure; another option, in development, provides for company infrastructure and customer keys. According to the company, automated systems can send limited safety signals without exposing prompts or responses to staff; the absence of broad human access does not prevent analysis. Customer storage can be durable. The implementation guide assigns the customer responsibility for retaining encrypted records for at least 30 days. The guide has no identifiable date; confirm the version and account conditions. The buyer should define the location, who is responsible for configuration and deletion, key control, automated access, and failure handling. Encryption does not eliminate processing or storage responsibilities.
What happened and what does the announcement support?
On November 16, 2026, OpenAI announced a classifier that produces scores to help prioritize review of images with possible CSAM, including AI-generated material. The company said it developed the classifier with support from NCMEC and offers it through an API at no cost to verified organizations. The announcement does not confirm deployment in the AI gateway.
OpenAI presents the classifier as the first widely available one with a specific signal for possible AI-generated CSAM. The claim to be first has not been independently verified. The announcement does not prove installation in an AI gateway or which modalities are analyzed by route. NCMEC recommends hashes and classifiers in services that store, transmit, or create images and videos, including generative AI. The organization says adoption is voluntary, varies among platforms, and that no technique is infallible. A recommendation is not proof of deployment and does not replace analysis of applicable law. The buyer needs to compare the announcement with the actual configuration of each contracted inference route. The LLM gateway evaluation guide organizes the broader selection process; here, the question is which data each layer processes and under what conditions it can preserve the data.
How does CSAM screening work in an AI gateway?
A flag does not confirm abuse. It is the result of CSAM screening in an AI gateway, not a report to authorities. Cryptographic or perceptual hashes depend on reference material; classifiers can identify content that has not yet been recognized. Review and reporting are separate steps, and an automated result can be wrong.
The AI gateway buyer should ask which service receives the CSAM flag, who examines it, what minimum data accompanies it, and how access is restricted. The buyer also needs to know the deadline or closure criteria for review, whether there is a channel for contesting a result, and its limits. Customer notification depends on the applicable conditions; it is not always permitted. The team can request configuration evidence and test the flow with lawful, safe data, never with abuse material. Without documentation, the answer is "not verified," not "does not exist." If internal policy requires review and it cannot be demonstrated, the route should be restricted for data subject to that control. Automated CSAM screening does not resolve questions about access, purpose, and subsequent retention. The article on endpoint evaluation and routing policy addresses how to measure route behavior.
Who reports, and when does the preservation period apply?
For possible CSAM in an AI gateway, in the United States, 18 U.S.C. § 2258A provides for covered providers to report to CyberTipline when they obtain actual knowledge of facts indicating the specified apparent violations. The law does not turn every automated flag into a mandatory report and does not impose an active content search under that section.
In cases of possible CSAM linked to an AI gateway, section 2258E(6) defines a provider, for sections 2258A through 2258E, as an electronic communication service provider or a remote computing service provider; whether an entity falls within that definition depends on the entity and its operation. section 2258A treats a completed report to CyberTipline under § 2258A(a)(1) as a request to preserve the reported content for one year after submission, pursuant to § 2258A(h)(1). § 2258A(h)(2) includes reasonably accessible contextual material, and § 2258A(h)(5) permits longer voluntary preservation under the conditions provided. The company should submit the flow for relevant legal review, without automatically transferring the obligation among the buyer, intermediary, and model provider. A contract does not change the law. It can establish duties between the parties. This description of United States law is informative, does not constitute legal advice, and does not establish an automatic obligation for the Brazilian reader: application requires analysis of the entity, operation, and jurisdiction.
How should inference data retention be compared by provider?
The useful comparison is not between a supposedly safe provider and another without safety measures. It is between a generic promise of zero retention in an AI gateway, without verified exceptions, and a documented commitment for each layer of the route, valid only under the conditions supported by evidence. The table below organizes procurement questions; it does not describe a product's configuration.
Retention in an AI gateway is verified by route, provider, and contracted configuration. The inventory should compare the planned route with the route actually executed by project, model, modality, and endpoint, including attempts, alternate routing, and relevant subprocessors. The buyer can request evidence of the configuration and records of the destination and conditions of the call, without sensitive content. In failures or when no eligible route is available, the team tests stopping the request when internal policy requires it. The test includes retries. Alternate routing, or fallback, receives the data only if it meets approved requirements or there is a permitted and formally accepted risk exception. The test identifies who authorizes and records the exception. The inventory covers transient memory, persisted copies, prompt and response caches, application state, uploaded files, error queues, technical traces, backups, and support or diagnostic data. For each class, procurement asks for location, access, the period or deletion criteria and when the period begins, including backup deletion and the handling of remaining copies.
| Stage or data | Question for the provider | Evidence required | Limit of what can be stated |
|---|---|---|---|
| Declared generic promise | Does the written promise cover the executed route, with exceptions, a time period, and a responsible party? | Contractual annex with version, date, and scope by route. | Without a documented exception, the item remains unverified. |
| Processing in transit | Is there screening? Which modalities and services participate? | Technical policy by service, model, and actual route. | Zero retention afterward does not prove the absence of analysis during processing. |
| Flagged content | When do retention, review, or access occur, and for how long? | Exceptions, purpose, responsible parties, access policy, and review period. | A flag does not confirm CSAM or, by itself, impose the one-year period. |
| Reported content | Who reports and preserves it, and under what conditions? | Applicable procedure, responsible entity, jurisdiction, and legal basis. | The obligation depends on the entity, facts, and applicable law. |
| Logs | Do records include content, identifiers, or only metadata? | Logging policy, recorded fields, access, and retention periods. | Do not assume logs and prompts receive the same treatment. |
| Retention by provider | Who receives data on each route, and which terms apply? | List of providers and subprocessors, versioned terms, and update procedure. | The intermediary's commitment does not prove the next provider's policy. |
Retention and training are separate questions: not using data for training does not prove the absence of storage, human access, or processing in another layer. The OpenAI data guide says that ineligible endpoints or features may retain application state even with ZDR enabled. This rule belongs to that provider; request evidence for the endpoint, feature, subservices, and configuration of the contracted account.
The AI gateway production index examines traffic mix and cost per request. For retention, the same route map has another use: linking each flow to its provider and the version of the terms governing it.
What should an AI gateway contract require before signing?
The AI gateway buyer should gather answers for each route before turning zero retention into an internal promise. The five actions below distribute the work among technology, security, privacy, legal, and procurement. None of them, on its own, demonstrates that a product already has a particular policy: each answer needs supporting evidence.
-
Identify entities and flows. Record the contracting party, intermediary, and AI gateway providers for each planned and actual route, including attempts and fallback. Test blocking when policy requires it.
-
Negotiate commitments by route. Check the contract and annexes for version and date; define whether content and metadata are covered, the services and routes, purposes, exceptions, deletion, and when the period begins. Establish the order of precedence among the contract, annex, and public policy.
-
Detail exceptions and review. For each exception, ask about the event, purpose, data, responsible party, access, deadline or documented completion criteria, and possibility of notification. Ask separately about flags, legal preservation, incidents, abuse, support, failures, ineligible features, and files.
-
Clarify reporting and preservation. Distinguish the contract, the law, and third-party policy; identify the responsible party, procedure, access, and notification limits.
-
Submit to legal and security. Request review of route evidence, applicable terms, and the handling of each exception, with a record of each decision.
The AI gateway buyer should negotiate the intermediary's commitment for its contracted chain, proportionate verification, and treatment of noncompliance. Security and technology check the route; legal assesses the entity and jurisdiction. A list of subprocessors does not demonstrate compliance. A material change in provider, route, subprocessor, or policy requires reassessment. The buyer should negotiate advance notice where possible, versioned evidence, an evaluation period and a responsible party, as well as handling of emergencies and objections. Until acceptance, the terms may provide for route deactivation, migration, or termination. There is no presumed universal right to veto or terminate. The same discipline appears in the continuity playbook for model deprecation. The difference is that, in this case, the test does not end when another model responds: the new route must meet the accepted data handling requirements before information is transferred through the alternative.
Frequently asked questions
Zero retention in an AI gateway, a flag for possible CSAM, and a report answer different questions: what remains after processing, what a tool flags, and when a responsible party submits a report. The answers below are limited to the policies and legislation cited; they do not attribute these procedures to a specific gateway.
Does zero retention prohibit screening?
Not necessarily. Zero retention in an AI gateway can address storage after processing, while screening takes place during the request. OpenAI documents this distinction in its offering, with an exception for images flagged as possible CSAM. For another route, consult the applicable terms.
Does an automated flag confirm CSAM?
No. A CSAM flag in an AI gateway can be mistaken and does not, by itself, confirm that the content is CSAM or constitute a report to authorities. NCMEC says that no solution is infallible and notes that critics point to the risk of false positives. Verification depends on the route.
Who may be required to report in the United States?
In the United States, electronic communication or remote computing providers may have to report possible CSAM in an AI gateway to CyberTipline when they obtain actual knowledge of facts indicating specified apparent violations under § 2258A. § 2258E(6) defines provider for sections 2258A through 2258E. Whether an entity is covered depends on the entity and operation, not merely on its presence in the chain.
Does the one-year period apply to every flag?
No. For possible CSAM in an AI gateway in the United States, § 2258A(h)(1) links the one-year period to a completed report to CyberTipline under § 2258A(a)(1), not to every flag or false positive. The law addresses reasonably accessible contextual material and permits longer voluntary preservation under the conditions of § 2258A(h)(5).
What should an intermediary be required to provide in an AI gateway contract?
The AI gateway buyer should negotiate commitments for each route and for the intermediary's contracted chain, covering data, purposes, exceptions, deletion, and means of verification. The contract and annexes need to clarify their order of precedence. Security tests the actual route, and legal assesses the terms for the operation before signing.
References and further reading
The AI gateway sources allow readers to check the zero retention policy announced by OpenAI, NCMEC recommendations on CSAM screening, and obligations under United States law, respecting the scope of each document and the law applicable to the specific operation. The version matters.
- OpenAI: classifier and child protection initiatives, November 16, 2026; OpenAI: zero retention and Private Safety Processing, August 19, 2026, updated September 22.
- 18 U.S.C. § 2258A: reporting and preservation; 18 U.S.C. § 2258E: definitions.
- NCMEC: minimum child safety measures for online platforms.
What commitment belongs on legal's desk?
Legal needs a record of outstanding items by route, with a responsible party and decision for each answer still missing. Nexforce Router is an AI gateway with multiple models and a unified API; its specific retention, screening, and review policies need to be requested and verified. The API does not prove them.
To evaluate Nexforce Router, procurement records documents, versions, dates, and evidence; security and technology check the configuration and the executed route; legal assesses the commitment and applicable law. Without an answer, the control remains "not verified," without assuming either presence or absence. The outstanding-items record turns that gap into a decision by route and data class. Approval of a route does not automatically approve fallback: the alternative requires the same analysis or an exception accepted under internal policy. After a material change, the responsible party checks the actual route, its terms, and the evidence again. Until the change is accepted, the negotiated consequence compatible with internal policy applies. The terms should specify scope, exceptions, and responsible parties. These criteria do not describe controls already verified in Nexforce Router and do not replace legal analysis.
| Recommended decision | Criteria and consequence |
|---|---|
| Approve | Written commitments and technical evidence cover the data, the actual route and its alternatives, and the exceptions are accepted. Approval applies only to the scope supported by evidence. |
| Restrict or stop | A mandatory internal policy requirement has not been demonstrated or an unapproved route may receive the data. Block the data subject to that requirement or stop the request until correction. |
| Record an exception | Only if internal policy permits acceptance of that risk: record the gap, scope of data and routes, authorized responsible party, justification, compensating measures, and reassessment deadline. Without that authorization, maintain the restriction. |
To examine the gateway and its routes, explore the Nexforce Router and request the retention, screening, and review terms applicable to the contract.

Save up to 50% in creditswith a single smart API
Connect your operations to our AI Router and optimize the consumption of multiple LLMs
Free TrialRelated articles

Tokens per second: what changed in LLM comparison
Artificial Analysis changed the base of its output speed methodology, which now includes time to first token and altered other aggregation criteria. Comparing sources requires declaring the denominator and measuring latency, percentiles and quality on your own load.
Read more
AI model deprecation: how to keep production running
A practical process for finding dependencies, testing replacements, and verifying traffic removal before the provider's deadline.
Read more
Price per million tokens: 50% less for input in 2026
A comparison across generations shows why routing decisions require task cost and quality, not token prices alone.
Read more