What changed, in one paragraph
On 27 October 2026 OpenRouter announced that it will screen all inference requests for child sexual abuse material (CSAM) as they pass through its service. Screening happens in flight, and content flagged as potential CSAM may be retained for human review, with reporting and preservation where the law requires it.1 The announcement says this also applies to requests covered by OpenRouter's Zero Data Retention (ZDR) policy.1 Its ZDR page carries a banner calling this a planned update and says the page and the related terms are being updated.2
None of this is unusual for a service that handles user content, and the stated purpose is a legal one. For a security reviewer the question is narrower: what does it change in the data-flow diagram, in the retention answer you give your own customers, and in the evidence you can collect?
What the published pages actually say
Read together, the pages give a layered picture. Each layer is a separate answer on a questionnaire.
- Default retention. OpenRouter says it does not store prompts or responses unless you opt in to logging or content is flagged as potential CSAM.3
- Opt-in logging. Private input and output logging is off by default. Org admins can view all logged data for an organization; members can view their own.3
- Opt-in use for product improvement. Off by default, offered in return for a 1% reduction in model usage cost.3
- Sampled categorization. OpenRouter samples a small number of prompts for categorization to power reporting and model ranking. Without the product-improvement opt-in, the categorization is stored anonymously and done by a model with a zero-data-retention policy.3
- Metadata. Token counts, latency and similar request facts are stored for every request. The page says this excludes prompt and response content.3
- Screening. In flight, on all inference requests, with the retention exception above.1
The ZDR page adds the exception to the headline. ZDR is defined as no routine storage after processing, "apart from the CSAM exception".2 A reviewer who has a "vendor offers ZDR" line item should now read it as "ZDR with a named, legally driven exception".
What screening covers, as described
The announcement describes two mechanisms. Known-material detection compares content against hashes of previously identified CSAM. Classifiers can flag suspected material not yet in a hash database, and OpenRouter says it may retain content flagged as potential CSAM for human review, including content that does not match a known hash.1 It also states that a flag indicates suspected material and is not a determination.1
On the legal side, it cites US reporting to NCMEC's CyberTipline and a one-year preservation requirement for reported material, and says that requirement applies to reported material rather than every item flagged for review.1
Data path: two inspecting parties instead of one
With direct provider calls, your prompt reaches one party under one contract. Through a router, it reaches the router and then a model provider, and the router's terms bind you to the provider's terms as well. OpenRouter's terms say you agree to comply with each model's applicable terms, are solely responsible for reviewing them, and that continued use after a provider changes its terms counts as acceptance.4 They also say that OpenRouter may suspend access to a model if a provider requests it, and that availability is "as-available" with no guarantee.4
For a reviewer, the screening announcement adds a content-inspecting step at the router. It also states that providers may run their own screening: some may retain content for a period to run automated safety checks before anything is flagged, under their own policies.1 So a single request can be inspected at two places, by two parties, for two sets of reasons.
Retention tiers you need to map
OpenRouter describes each endpoint with three data policy tiers.2
| Tier | What it means | Trained on? |
|---|---|---|
| ZDR | No routine storage after processing, with the CSAM exception | No |
| Safety Retention | OpenRouter or a provider may retain prompts and responses from any request, flagged or not, only to detect and prevent abuse | No |
| Retention and Training | Stored, and may be used for training under the provider's policy | May be |
Three details matter in a review.
- Safety Retention is not limited to flagged requests. The ZDR page says it limits why data can be kept, not whether a request has to be flagged first.2
- Self-serve defaults differ from enterprise defaults. If you have not chosen, self-serve accounts allow Safety Retention endpoints and enterprise accounts use ZDR only.2 Ask which one your account is on and where the setting is recorded.
- Unclear policy is treated conservatively. When OpenRouter cannot establish a provider's or endpoint's policy, it assumes the endpoint both retains and trains and marks it that way.2
Also note that ZDR enforcement applies to provider routing for inference. It does not cover plugins and tools you enable, such as web search, which may be run by third parties with their own retention policies.2 And OpenRouter says it has no routing rules that change on provider retention policies by themselves; you or your admin set the filters.5
Evidence: what you can and cannot hand an auditor
A security review ends in artifacts. For a router in the data path, sort them by who produces them.
| Question | Evidence source | Status from the pages read |
|---|---|---|
| Which endpoint served a request, under which data tier? | Router logs and endpoint policy data | Metadata is stored per request;3 the tier is a published property of the endpoint2 |
| Was my content retained by the router? | Router | Opt-in logging is visible to admins;3 retention of flagged content is not described as visible |
| Did screening flag my traffic? | Router | Not described |
| What did the provider retain? | Provider policy and your agreement | Published per provider, as the provider's own terms5 |
| Where was it processed? | Regional hostname | In-region routing for EU or US exists on Business and Enterprise plans, selected by sending requests to the regional hostname5 |
The gaps in the third and fourth rows are the ones to write down. They do not mean the vendor will refuse; they mean the pages do not commit to an answer.
How a BYOK gateway with a direct contract differs
Be exact about the contrast, because it is easy to overstate. A BYOK gateway is still in the data path. Your requests pass through it, so it belongs on the same diagram and the same questionnaire. What differs is where the contract and the account sit.
With INFRO, you connect provider accounts you already hold, and your providers invoice you directly; INFRO takes 0% of model spend and charges a subscription for the control plane.6 Your providers process content under your own agreements and account settings, and INFRO's privacy page says it does not train on or sell your content.7 That keeps three things yours: the provider's retention and ZDR terms as you negotiated them, the provider-side logs and audit trail, and the choice of which provider accounts exist at all.
Router-side retention is also a setting you own. INFRO's content logging is off until an owner or admin turns it on; it can be limited to prompts, responses or both, and a request can set logging: false. Payload retention is the smaller of your chosen window and the plan maximum.7 The same page states that disabling INFRO logging does not disable storage or logging at your AI provider,7 which is the sentence to keep in your own diagram: two stores, two settings.
OpenRouter also supports BYOK, and a reviewer should check how it behaves. Requests on your keys still go through OpenRouter, BYOK usage carries a fee beyond a plan-dependent free allowance, and by default a failure on your key can fall back to OpenRouter's shared endpoints. Setting the fallback to "never use shared capacity" keeps requests on your account.8 If your reason for BYOK is a direct provider contract, that fallback setting decides whether the contract actually covers all traffic.
Controls that stay on your side of a gateway
These INFRO capabilities are available now and are relevant to a data-path review:
- Provider connections. Credentials are encrypted when saved, scoped to the organization and never shown again; applications use gateway keys instead.
- Guardrails. Pattern-based checks on prompts and responses (email addresses, card numbers, API keys, your own terms) that block, redact or flag. Detection can miss things, so it is a control, not a guarantee.
- Model allowlists and key limits. Approve which models a project may call, and restrict a gateway key to chosen providers or IP ranges.
- Audit log and OpenTelemetry export. Member, key and policy changes with actor and timestamp, and spans pushed to your own collector.
Some things a reviewer may ask for are not available. Regional processing, a staff access log, security event export to a SIEM, and self-hosted deployment are not shipped; treat them as roadmap or enterprise-agreement items. INFRO does not hold SOC 2 or ISO certifications. The full API contract is published in the INFRO docs (opens in a new tab) for review before any code is written, and the privacy and data page (opens in a new tab) covers logging and retention in more detail.
What to tell your security reviewer
Give the reviewer a short, dated statement rather than a link to marketing. A workable version:
Then work through this list before sign-off:
- Redraw the data-flow diagram with the router as a content-inspecting party, plus the model provider behind it.
- Record the terms version and date you accepted, and re-check after the announced documentation and terms update lands.
- Confirm the account's Safety Retention setting and who can change it.
- List every plugin and tool enabled, since ZDR enforcement does not cover them.2
- Update customer-facing ZDR language so it names the screening exception where you repeat the vendor's claim.
- Ask in writing how a customer is notified of retained content, and how long review retention lasts.
- If you rely on BYOK, check the shared-capacity fallback setting on every key.
- Decide which workloads, if any, must use a direct provider contract with no intermediary content inspection beyond your own gateway.
The last item is the real decision. Screening is a legitimate, legally anchored measure, and it does not make a router unsafe. It makes the router a party that inspects content, and that belongs in your risk record next to the provider's own retention terms.
