Golden Secrets / Secure Input
The safest credential is one the AI never needs to see. Golden Terminal is being designed to let an AI use a password, token or other secret for an approved action without putting the raw value into the model conversation. Until Secure Input ships, keep secrets on the connected machine whenever possible and give the AI a reference, not the value.
Quick answer: how should I give AI access safely?
Let the connected computer hold the credential. Let the AI know how to reference it.
For SSH, keep the private key in your local SSH configuration. For APIs, prefer an existing CLI login, operating-system credential store, environment variable or protected local secret file. In Canon or project instructions, store the alias, variable name or file path — not the secret value.
AI knows which credential to use. Your computer knows the credential value.
What works today
Golden Terminal executes actions on the currently connected device in the permissions of the current user. That device can use credentials already configured in its local environment without requiring you to paste the raw value into the conversation.
Examples include SSH profiles, authenticated developer CLIs, macOS Keychain, Windows Credential Manager, environment variables and protected configuration files.
Do not ask the AI to print a secret file just to make the credential available. A command that reveals a secret can place that value in command output and therefore into the AI conversation. Secret redaction is a safety layer, not a substitute for keeping the value out of output.
Recommended credential methods
| Method | What you tell the AI | Where the secret stays | Use |
|---|---|---|---|
| SSH key + SSH config | Server alias, for example production | Private key on the connected computer | Recommended |
| Existing CLI or OAuth session | Use the existing authenticated session | Local session or credential cache | Recommended |
| OS credential store | Credential name or reference | Keychain / Credential Manager / local vault | Recommended |
| Environment variable | Variable name such as PROD_DB_PASSWORD | Connected computer | Good |
| Protected local secret file | File path and how it should be consumed | Connected computer | Good |
| Canon reference | Alias, variable, file or credential-store reference | Separate local storage | Good |
| Temporary scoped token | Raw token only if there is no local alternative | AI conversation during use | Fallback |
| Permanent password or unrestricted token | Raw secret | AI conversation | Avoid |
Canon is a map, not a secret vault
Canon and project instructions are useful for remembering where access is configured. They should contain references to secrets rather than secret values.
Good
Avoid
The same rule applies to notes, README files and prompts that are routinely copied into an AI conversation. Keep durable instructions useful while keeping the sensitive value somewhere designed to protect it.
Local files and environment variables
A local secret file can be useful when the command consumes it without printing it. Examples include a protected .env file, a service configuration file or an application-specific credentials file.
Prefer restrictive file permissions and application-specific access. If a secret can be injected directly into the process that needs it, avoid intermediate commands that display it.
If you must provide a token directly
Sometimes a local credential is not available. In that case, prefer a temporary token over a permanent password.
- Give it only the permissions required for the task.
- Restrict it to the smallest possible resource or project.
- Use a short expiration time.
- Revoke or rotate it after the task.
- Do not reuse a temporary troubleshooting credential as a permanent production credential.
A temporary token pasted into an AI chat is still present in that conversation context. Short lifetime and limited scope reduce the impact; they do not turn the conversation into a password manager.
What should not be pasted into AI
Avoid sending long-lived or recovery-grade secrets when another workflow is possible.
- SSH private keys
- Seed phrases and cryptocurrency private keys
- Master passwords
- Account recovery codes
- Permanent administrator passwords
- Long-lived unrestricted API tokens
- Full secret files when only one local action needs them
If a credential has already been exposed in a conversation or command output, treat rotation or revocation as the safest response.
Golden Secrets / Secure Input: planned flow
Golden Secrets is in development. The target architecture is a secure path for one-time or short-lived credentials where the raw value bypasses the AI model whenever possible.
The model identifies that an approved action needs a password, token or other secret.
The user enters the value in a dedicated Golden interface instead of typing it into the AI conversation.
The AI receives a handle such as secretRef, not the raw secret.
The credential is supplied only to the approved local operation through the narrowest suitable mechanism.
The model gets operation status and sanitized output without needing the credential value.
The handle is destroyed or becomes unusable after the approved task or timeout.
Design goals
- Keep raw secrets out of the model prompt and normal conversation context.
- Avoid raw secret values in tool arguments and tool results.
- Avoid writing secret values to normal application logs.
- Use explicit scope, short lifetime and one-task semantics.
- Let the user see when a protected credential is requested and where it will be used.
- Audit the action without recording the secret itself.
The example above describes the product direction. It is not a public API contract and should not be treated as a currently available Golden Terminal capability.
Current status: in development
Golden Secrets / Secure Input is not available yet.
Do not paste a password into ChatGPT and expect Golden Terminal to hide it automatically. Until Secure Input is released, use the local-first methods on this page: SSH keys, existing sessions, credential stores, environment variables, protected local files and short-lived scoped tokens when a direct token is unavoidable.
FAQ
Does Golden Terminal need my SSH private key?
No. Keep the private key on the connected computer and configure SSH locally. Ask the AI to use the SSH profile or host alias instead of sending the key itself.
Can I store a password in Canon?
Do not store the raw password. Store a reference such as an environment-variable name, SSH profile, file path or credential-store entry.
Can I upload a .env file to the AI?
Prefer keeping the file on the connected computer and telling the AI where it is or which variables should be consumed. Uploading the full file exposes every value in it to the conversation context.
Is automatic secret redaction enough?
No. Redaction can reduce accidental disclosure of recognizable patterns, but arbitrary command output may contain values that no filter can reliably identify. The safer pattern is to avoid printing the secret in the first place.
When will Golden Secrets be available?
It is in development. There is no public launch date yet. This Help page will be updated when Secure Input becomes available.
Keep access local
Configure the credential on the computer that Golden Terminal can reach, then ask your AI to use the local profile, variable or credential reference.