Data Protection

Handle a secret caught in chat, read the placeholders agents show in place of secrets, and know what the model sees.

CloudThinker catches secrets on the way in, when you paste them into chat, and on the way out, when an agent’s tools return them. This page shows what you’ll see in each case and what to do about it.

Prerequisites

• The paste scan and exact-match masking of secrets CloudThinker holds are always on, for everyone who can chat. There’s nothing to turn on.
• Masking of secret formats you never saved follows the API keys and Cloud credentials rows in Guardrails. Both are on by default, but a workspace Admin can turn them Off unless an organization Admin locks them.
• To save a caught secret as a credential: the Admin workspace role. Developers can still redact or keep the text.

What’s protected, and where

Moment	What CloudThinker does	What you see
You send a message, or paste long text	Scans it for credential-shaped values such as API keys, tokens, passwords, and connection strings	The Review detected secrets dialog, before anything sends
An agent’s tool returns a secret	Replaces the value before the model, the chat, or stored output gets it. It checks two ways: an exact match against secrets CloudThinker holds for the run (connection secrets, and workspace credentials saved under a secret-sounding name such as API_KEY, TOKEN or PASSWORD), and the API keys and Cloud credentials pattern rules in Guardrails. Those two rows are on by default but not locked, so a workspace Admin can turn them off. While they’re on, an AWS secret key in a config file an agent prints is hidden even if you never saved it.	A placeholder such as ct:secret:v3:openai-api-key:2d9c4e7a0b16
An agent’s tool returns a protected identifier, such as an email address	Replaces it the same way, using your Guardrails settings	A placeholder such as ct:pii:v3:email:4c1f0a9e27b3 [op***@ma***.com]

What the model sees

The model gets your message after the paste scan, and agent tool output after placeholders replace secrets and protected identifiers. It never gets the real value behind a placeholder.

Handle a secret you pasted

Say you paste a message that includes a live key:

Check why this key gets 401s from the billing API: sk-live-...

1. Send the message

   Press Enter as usual. The message doesn’t send. The Review detected secrets dialog opens and lists each value it found.

2. Check each item

   For a false positive, such as an example string, click Mark as not a secret on its row. That text stays as you wrote it, and it sends as typed, so only mark text you’re sure isn’t a real secret. If you’re an Admin, each row also shows the credential key name it will be saved under. Edit it if needed.

3. Choose what happens to the rest

   Button	Result
   Save secrets (Admin only)	Stores each value as a workspace credential under its key name, and puts the key name in your message where the value was
   Redact from message	Replaces each value with [sensitive data removed]
   Keep in message	Shows instead of Save secrets when you’ve marked every item as not a secret. Sends the text unchanged, real secret or not.
   Close the dialog (Esc)	Nothing sends. A typed message stays in the box with the value still flagged, so you can edit it. A long paste is discarded.

   With Save secrets, the example message sends as:

   Check why this key gets 401s from the billing API: $BILLING_API_KEY
   

   A command that uses $BILLING_API_KEY gets the real value when it runs. The model and the chat only see the name.

   Success state: the message sends with a $KEY name or [sensitive data removed] where each flagged value was. An item you leave flagged can only be saved or redacted. An item you mark as not a secret sends as typed, so marking a real secret that way puts it in the conversation.

The scan covers the text you type and the text you paste, including a long paste that becomes an attachment. Next time, skip the dialog by referencing a saved key. Workspace Admins can type $ in the chat box and pick it from the list. Other members type the name in full, such as $BILLING_API_KEY. See Use a credential in chat.

Read a placeholder in agent output

When an agent’s work would show a secret (one CloudThinker holds, or one that matches a Guardrails credential rule) or a protected identifier, you see a placeholder instead.

• The same value gets the same placeholder within a conversation, so you and the agent can still tell that two references point to one secret. In another conversation, the same value gets a different placeholder.
• A placeholder can’t be turned back into the value, by you or by an agent. There’s no reveal button.
• The parts of a placeholder: ct, then secret or pii, the format version (v3), the category (such as openai-api-key), and an ID. The ID is a keyed HMAC-SHA256 value that mixes in the conversation, so it can’t be brute-forced back to a short password without CloudThinker’s key.
• Identifier placeholders may carry a hint in brackets, such as [***-***-**89] for a phone number, so the output stays readable.

To change the secret itself, update it in Connections or Credentials. To choose which identifiers are hidden, use Custom Guardrails.

If you open a CloudThinker demo replay, secrets and identifiers show as ***REDACTED***. Customer workspaces can’t create replay links yet. See Sharing.

Unsent drafts

Text you’ve typed but not sent stays in your browser, not on CloudThinker’s servers, so a reload doesn’t lose it. A conversation draft is kept for up to 24 hours. Logging out clears every saved draft.

The secret scan runs when you send, not while you type. On a shared computer, log out instead of closing the tab if a draft might hold a secret.

Troubleshooting

The dialog flagged something that isn't a secret

Cause: the value looks like a key, for example a long random test string.

Fix: click Mark as not a secret on that item. When every item is marked, click Keep in message and the message sends unchanged.

I don't see Save secrets

Cause: saving creates a shared workspace credential, which needs the Admin workspace role.

Fix: choose Redact from message, then ask a workspace Admin to add the key in More → Workspace settings → Credentials. See Credentials.

An agent shows a placeholder where I need the real value

Cause: the value matches a saved secret or a protected identifier.

Fix: for a secret, reference it as $KEY so commands get the value without it showing in chat. For an identifier, a workspace Admin can turn that protection Off in Admin Settings → Guardrails, unless an organization Admin has locked it.

A secret got through the scan

Cause: the scan looks for known credential formats. An unusual format can be missed.

Fix:

1. Rotate the secret at its provider. Assume the model saw it.
2. Ask a workspace Admin to delete the conversation (Delete Conversation in the conversation’s title menu), so the raw value isn’t kept in its history.
3. Save the new value as a credential and reference it as $KEY.

To catch that format next time, a workspace Admin can add a custom pattern. It masks matches in tool output (as a ct:pii placeholder), not in the paste scan. For a secret format the scan should know, tell CloudThinker support.

Next steps

Custom Guardrails

Choose which identifiers agents see as placeholders

Credentials

Save secrets once and reference them as $KEY