Cache-Control no-store Policy for LLM Prompts and Responses
Cache-Control no-store tells compliant private and shared caches to avoid retaining an LLM request or response. This guide applies RFC 9111 to authenticated inference traffic, separates no-store from no-cache and private, defines placement at each HTTP hop, and specifies release tests and audit evidence without presenting a cache directive as a complete privacy control.

An authenticated inference request can cross a browser cache, service worker, forward proxy, and reverse proxy before passing through an API gateway and provider adapter on its way to a model. Each cache is a possible copy point for prompt text and retrieved context, plus tool results and model output. Cache-Control: no-store gives those HTTP components a direct storage instruction. It needs to appear on the correct messages and survive every hop.
TL;DR
- Send
Cache-Control: no-storeon sensitive LLM requests and responses. - Apply the directive across private and shared caches, then verify each deployed hop.
- Keep
no-store,no-cache, andprivateseparate because they produce different storage and reuse behavior. - Treat the directive as one cache control rather than a complete privacy guarantee.
- Record the effective policy and observed headers without copying raw prompts into logs.
no-store blocks intentional cache storage
RFC 9111 Section 5.2.1.5 says the request directive prevents a cache from storing any part of the request or its response. It applies to private and shared caches. The defined behavior bars intentional non-volatile storage and calls for a best-effort removal from volatile storage after forwarding.
The response form carries a related but distinct instruction. RFC 9111 Section 5.2.2.5 tells a cache to store no part of the immediate request or response and bars reuse of that response for another request. An LLM API that controls both directions should set the request directive before forwarding and attach the response directive before any downstream cache can handle the output.
I consider a sensitive inference route without an explicit cache policy unfinished. Authentication limits who can call the route. It says nothing about copies left in compliant HTTP caches after the call.
The standard also sets a hard limit on the claim. no-store is insufficient as a privacy mechanism because a malicious or compromised cache may ignore it and a network may permit eavesdropping. Transport security, cache configuration, endpoint controls, retention policy, and access control keep their own jobs. The AI egress control implementation guide places this cache instruction inside the wider outbound path.
no-cache and private solve different problems
The directive names invite mistakes because, under RFC 9111 Section 5.2.2.4, an unqualified no-cache response can be stored, but a cache must validate it successfully before using it to satisfy another request. That creates a reuse rule, while no-store supplies the storage rule for a prompt route that should leave no intentional cache copy.
The private response directive prevents storage in a shared cache while allowing a private cache to store the response under RFC 9111. That can suit a user-specific web page. It is a loose default for prompts or outputs containing enterprise data because a browser or other private cache remains allowed to retain them.
must-revalidate also permits storage. It controls reuse after a stored response becomes stale, and Section 3.5 lists it among the directives that can allow a shared cache to store a response to a request carrying Authorization. An authenticated request therefore deserves an explicit route policy. The mere presence of an authorization field does not express the organization's intended retention rule.
A practical sensitive-route response can carry Cache-Control: no-store. Teams sometimes add private or no-cache for old implementation behavior, but those extra tokens cannot strengthen a component that correctly implements no-store. Document any compatibility addition and test the actual cache chain instead of treating a long header as stronger policy.
Put the directive on both sides of the routed call
Set request no-store at the first trusted component that has route and identity context, before a forward proxy or shared gateway can store the exchange. Preserve it on the outbound request to the model endpoint. On the return path, the application or gateway should add response no-store before the bytes reach a downstream reverse proxy, browser, agent client, or service worker.
The route manifest should state which calls require the directive. Chat completions, embeddings generated from private text, retrieval-augmented prompts, and tool-bearing model requests commonly carry sensitive material. A policy can also cover error bodies because validation traces or rejected prompt fragments may appear there. Apply the rule by authenticated route and data class instead of relying on developers to remember a header in each handler.
Draw the deployed chain on a whiteboard. Put a red square around every component with cache capability, then mark the request and response header observed at its ingress. This physical map catches an awkward fact that an application unit test misses: a content delivery layer or service worker can make a storage decision before the header is added.
The identity-aware proxy for LLMs explains how verified caller context can select the route policy. The cache decision still belongs to the HTTP exchange and should remain visible as a separate policy outcome.
Existing entries and transformations need explicit tests
A request carrying no-store can still be satisfied by a response that was already stored. RFC 9111 states that the request directive does not apply retroactively to such an entry. A rollout should purge existing cache entries for affected routes, disable cache lookup where the platform supports it, and verify that a previously stored object cannot satisfy the authenticated request.
Test every intermediary with a unique synthetic marker rather than real prompt content. Send the marker through the authenticated LLM route, inspect cache status fields and protected cache administration views, then repeat the request after disconnecting the origin in a controlled environment. The second request must fail or reach the origin according to the test design. A cache hit exposes a policy or configuration gap.
Include successful outputs, streaming outputs, client errors, gateway errors, retries, and cancelled streams. Confirm that header normalization preserves no-store; a proxy must not replace it with private, no-cache, or a positive freshness lifetime. The LLM security testing guide provides the broader pattern for boundary fixtures and expected decisions.
Audit evidence should prove policy and observation
For each routed decision, retain the authenticated subject or agent, normalized route, model destination, data classification, cache-policy version, request directive applied, response directive observed, decision, status, and correlation ID. Record the deployed intermediary inventory and test result as release evidence. Raw prompts and model responses make poor cache-policy logs because the evidence system would become another content-retention surface.
A configuration check and a wire observation answer different questions. The route manifest shows intended policy. A capture at the gateway boundary shows the header that crossed that point. Cache administration evidence shows the resulting storage behavior. Keep all three tied to the same release and synthetic test identifier.
Direct calls outside the governed route retain separate cache behavior. So do local model execution and non-HTTP transports. The post-authentication gap is the relevant operating principle: a valid identity begins the request decision, while route-specific controls determine what happens to that request.
DeepInspect
DeepInspect can apply an identity-aware cache policy to authenticated user and agent traffic deliberately routed through its HTTP path to an LLM endpoint. At that boundary, it can evaluate caller context, route, model destination, and data class, then enforce the approved request and response handling decision before forwarding content.
The resulting audit record can capture the policy version, effective route, directive decision, status, and correlation context without retaining raw prompt text by default. Cache software, client-side storage, and direct provider traffic remain under their respective owners, so the release fixture still has to exercise the full deployed HTTP chain.
Book a demo today.