Skip to Content
Onshore is in early access. Don’t load real customer data until the items in Known limits are closed.
DevelopersTask guidesCloak text for AI

Cloak text for AI

AI models usually run outside your customers’ country. Cloaking replaces identifiers in a prompt with placeholders, stand-in names or less specific values before it leaves, and keeps a sealed map (a digital twin) in the vault so the model’s reply can be restored for the person who’s allowed to see it.

"Hi, I'm Sarah Johnson, 34. My card 4111 1111 1111 1111 was charged £1,250.50 twice." │ cloak (in the country) ▼ "Hi, I'm Emily Carter, 34. My card [CARD_1] was charged £1k–£2k twice." ──▶ model abroad ▲ │ uncloak (in the country, as access rules allow) "Sorry Sarah, I've flagged the duplicate charge on ****1111." ◀── "Sorry Emily, I've flagged … [CARD_1]."

Two ways to do it:

AI gatewayCloak API
HowPoint your OpenAI or Anthropic SDK’s base URL at the vaultCall cloak and uncloak yourself
Code changesOne lineA few calls around the model call
Works withOpenAI-compatible and Anthropic APIsAny model, any provider, any outside service
GuideAI gatewayThis page

Choose what each kind of data becomes

Data typeChoicesRestored by uncloak?
person_nameplaceholder → [NAME_1], synthetic → a realistic stand-in, keepPlaceholder and stand-in: yes
amountplaceholder, band → £1k–£2k (same currency), keepPlaceholder: yes. Band: no
date (dates of birth)placeholder, age_band → aged 35–44, year → 1988, keepPlaceholder: yes
addresskeep (default), state → New York, region → United States, placeholderPlaceholder: yes
phone, card_pan, payment_refplaceholder, last4 → ****1234Placeholder: yes
emailplaceholder, domain → a gmail.com addressPlaceholder: yes
phone_ngplaceholder, network → an MTN number, last4Placeholder: yes
Country pack types: uk_nino, iban, us_ssn, in_aadhaar, br_cpf, za_id, …placeholder, last4Placeholder: yes

Anything not listed, including your own data types, becomes a placeholder.

Which to pick: stand-in names make replies read naturally (“Dear Emily”) and models handle them better than [NAME_1]. Bands and age ranges keep what the model needs to reason (“a mid-sized refund”, “a customer in their thirties”) without the exact value. Use keep only when the model genuinely needs the value and your data protection assessment allows it.

Save a combination in the console as a cloak setting (Connections → AI → Cloak settings) and pass its profileId, so every service uses the same choices and admins can change them without a deploy. Or pass options inline.

Cloak, call the model, restore

import OpenAI from "openai"; const openai = new OpenAI(); export async function answer(chatId: string, customerMessage: string) { const cloaked = await vault.cloak(VAULT, customerMessage, { profileId: "cp_support_chat", keepDays: 1 }); const completion = await openai.chat.completions.create({ model: "gpt-5", messages: [ { role: "system", content: "You are a bank support assistant. Placeholders like [CARD_1] stand for real values; repeat them exactly." }, { role: "user", content: cloaked.text }, ], }); const restored = await vault.uncloak(VAULT, cloaked.id!, completion.choices[0].message.content ?? "", `Support chat ${chatId}`); return restored.text; // what the agent or customer sees }

Tell the model in the system prompt to repeat placeholders exactly. Models sometimes reformat [CARD_1] as CARD_1 or translate it, and then it can’t be restored.

Who sees what after uncloak

Uncloak is a reveal. Each kind of value comes back in full only if the caller’s access rules allow plaintext for it; otherwise it comes back masked. decisions tells you what happened:

const { text, decisions } = await vault.uncloak(VAULT, twinId, reply, `Support chat ${chatId}`); // decisions = { person_name: "plaintext", card_pan: "masked" }

So a support bot’s service account can restore names but leave card numbers masked, with no code changes, just access rules.

Multi-turn conversations

Cloak each new message, and keep the conversation history cloaked on your side: send the model the cloaked history, uncloak only what you show. Keep one twin per message (each cloak call returns its own id) and uncloak a reply with the twin of the message it answers. If a reply refers back to values from several earlier messages, uncloak it with each relevant twin in turn.

Placeholder numbering ([NAME_1]) starts again in each call, so [NAME_1] in message 1 and [NAME_1] in message 3 can be different people. Cloak the whole conversation in one call when the model must tell people apart across turns.

Streaming

Uncloak works on complete text: a placeholder split across two chunks ([CA + RD_1]) can’t be restored. Either buffer the full reply and uncloak once, or buffer up to the end of each sentence and uncloak sentence by sentence.

let buffer = ""; for await (const chunk of stream) { buffer += chunk.choices[0]?.delta?.content ?? ""; const end = buffer.lastIndexOf(". "); if (end > 0) { const ready = buffer.slice(0, end + 2); buffer = buffer.slice(end + 2); send((await vault.uncloak(VAULT, twinId, ready, purpose)).text); } } if (buffer) send((await vault.uncloak(VAULT, twinId, buffer, purpose)).text);

Tool calls

When the model calls a tool, its arguments contain placeholders (refund_card({"card": "[CARD_1]"})). Your tool runs on your side, so uncloak the arguments before running it, then cloak the tool’s result before sending it back:

const args = JSON.parse((await vault.uncloak(VAULT, twinId, call.function.arguments, `Tool ${call.function.name}`)).text); const result = await tools[call.function.name](args); const back = await vault.cloak(VAULT, JSON.stringify(result), { profileId: "cp_support_chat" }); messages.push({ role: "tool", tool_call_id: call.id, content: back.text });

The tool’s service account needs plaintext for the fields it uses.

Twins

  • keepDays from 1 to 365 keeps a twin; 0 (the default) keeps none, which is right for one-way uses such as sending text to a translation service or analytics.
  • Delete a twin when the conversation ends with deleteTwin; it expires on its own otherwise.
  • An expired or deleted twin fails uncloak with not_found.
  • Twins are sealed with the vault’s keys and never leave the country.

Limits

  • Text up to 256 KB per call (413 too_large).
  • Ranges, age bands, years, states and regions can’t be reversed.
  • Cloaking relies on detection, so an identifier in an unusual format can be missed. Try your real prompts in the Playground on the AI tab of Connections before going live, and add your own data types for your identifiers.

Next

Last updated on