AILE.SH
$AILE $0.0000223
EN
Log in

Who can read your prompts on Aile?

Privacy depends on the route your request takes. What sharers, Aile's relay, providers and agent operators can see, and why "blind" doesn't mean invisible to everyone.

Note 04Jason

Before you send a prompt, ask two questions. Which model will answer? And who can read what you send?

These are separate choices. Aile connects callers to several kinds of capacity through one API, but the routes don't have the same privacy. A machine that runs a model has different access from a machine that only forwards encrypted traffic.

One thing holds on every route: Aile reads your request and the response so it can route, check and meter the call. The routes differ in who else can read them.

Through a node: the node forwards, the provider answers

On a route through a node, the sharer's machine forwards encrypted traffic. It can't read your prompt or the provider credential.

Aile's relay prepares the authenticated request, so it can read the prompt and the response. The provider then processes the request to produce the answer. A sharer's node that can't read your prompt doesn't mean Aile can't.

"Can read" here means having access to readable content. It doesn't mean a person reviews every request.

Nodeless: no sharer machine in the path

On a nodeless route, Aile connects to the provider directly. No sharer node forwards the request, but Aile and the provider can still read it.

Taking the node out changes the route. It doesn't hide the prompt from the systems that handle it.

Self-hosted: the sharer runs the model

A self-hosted sharer runs the machine that does the inference. That machine reads the prompts it answers, because processing them is the job. Aile still reads the request to price and route it. No model provider sees it.

"Self-hosted" means the sharer's hardware, not your own device. Calling a self-hosted listing means trusting another person's machine with your input. The model weights stay on that machine, which is a separate question from who can read your prompt.

You can keep a key off self-hosted machines entirely. Untick Allow self-hosted model in the key's routing rules.

What "blind" actually means

A reply served through a node may include:

x-aile-capacity: subscription
x-aile-blind: true

The second header describes the sharer's node, not the whole path. It means that machine couldn't read your request. It says nothing about Aile's relay.

Nodeless replies say x-aile-blind: false and leave out x-aile-node, because no sharer machine took part. That describes the route. It doesn't mean another sharer received your prompt.

Use these headers to learn how a request was served. Don't treat one flag as an end-to-end privacy guarantee.

What Aile hides from sharers and providers

On the chat endpoints, Aile looks for secrets in your prompt before it leaves. That covers private keys, recovery phrases, API keys with a known prefix, and .env values. Each one is swapped for a placeholder, so the sharer and the provider never receive it. If the model writes a placeholder back, Aile puts the real value back before your client sees the answer.

This protects you from the sharer and the provider, not from Aile, which still reads the original. It also only catches secrets in shapes it recognises. It's on by default. x-aile-redact: off turns it off for one request.

Two more controls are worth knowing:

  • x-aile-data-collection: deny refuses any model that lacks a sourced no-training policy. Self-hosted capacity is included in that refusal.
  • x-aile-hidden-prompt: deny skips accounts measured adding a system prompt of their own. The measurement is evidence, not proof.

Reading is different from storing

Aile's Privacy Policy says request and response bodies are read in memory to route, screen and meter each call, and are not stored.

It lists billing metadata separately: timestamps, model names, token counts, latency and status. Operational logs are kept for 30 days. That 30-day period doesn't cover every kind of account or billing data.

Not storing a prompt doesn't mean Aile can't read it while handling it. Aile's storage policy also says nothing about each upstream provider or each sharer's own machine.

Encryption has limits too. Linked provider credentials are stored encrypted, but Aile decrypts them in memory to serve requests. Encrypted storage doesn't keep them out of Aile's reach.

Agent tools have another reader

Calling another agent's tool is different from calling a model.

Aile sends the task text and the arguments the sharer's tool declares. It doesn't add your conversation history, system prompt, memory or other tools' output.

But the task has to reach whoever does the work. Aile and the sharer's agent can read it. On an attended listing, the person who owns that machine can see it too. Don't send secrets or confidential documents to an attended task.

Your name, email, wallet and IP aren't passed to the sharer, and the sharer can't link two of your calls. That's not the same as hiding what you type, though. If you paste a customer's name, email or a private document into the task, the recipient gets it.

Send the smallest useful input

Say you need help rewriting a support reply. Send the relevant issue, and swap identifying details for placeholders. Don't include the customer's whole account record just because you have it.

Check what your app actually sends, including history and tool arguments, not just the last message you typed. Keep credentials and private keys out of prompts, and use public or made-up material while testing.

Before you send sensitive data, find out who runs the capacity you picked and what rules apply to that data. Words like "self-hosted", "encrypted" or "blind" don't mean nobody else can read your input.

Pick the route by who you trust with the input, not only by which model you want.