Secret Protection Guide & User Manual

Translation notice: This page is translated from Korean. You can also read the Korean source using the language menu. Your statutory rights remain unchanged regardless of language.

By default, we restrict AI tools from receiving raw secrets. This is not a product that prevents OS compromise or access through other tools, or guarantees secrecy in every circumstance.

Optional sharing of values: ‘Not secret’ is off by default. Once you confirm the warning and turn it on, approved AI connections can read the original value. This is separate from permission to use a private value. The AI service's processing and retention policies apply to values it receives.

Values already sent to AI cannot be recalled by turning this setting off later. Choose carefully before sharing. Turning it off blocks future reads of the original value, but cannot recall values in transit or already delivered. Editing a value preserves your sharing choice. Turn this setting off before entering a new secret. Restoring a backup as a new vault resets AI connections, access and sharing settings.

How to use it

Store secrets in value fields and allow their use separately with the AI access toggle and the approval card. AI sends data labels and task inputs; Bossanova checks approvals locally and places required values only into approved destinations or launched processes. AI receives limited receipts, statuses and masked results for each task. Destinations receive actual values, and labels/descriptions may be visible to AI.

Currently supported are two operations: “Send on my behalf” (Bossanova sends an HTTPS request to an approved service) and “Inject into a run” (Bossanova launches an approved executable and passes the value through a carrier). Filling in browser screens or generating contracts containing secrets is not a feature. The PC and Bossanova must be running, and secrets are not used while the vault is locked. Receipt does not mean completion.

Outside the protection boundary: prevention

• Secrets in names or descriptions: Use labels that are safe to share. Do not repeat secret values in titles or descriptions.
• Direct disclosure: AI can see secrets you paste into chat or provide in files, screens or exported results. Also consider clipboard history and synchronization after viewing or copying values.
• Excessive AI permissions: Administrator privileges, unrestricted shell commands, file/screen/accessibility/clipboard access, debugging and memory access are separate from Bossanova connection approval. Choose approval durations (especially ‘Always’) carefully and allow only what is needed. Disconnecting Bossanova does not revoke permissions granted to other tools.
• Compromised device: OS/account compromise, malware, keyloggers and modified apps can intercept values or passwords. Use official releases and security updates. Vault encryption does not make a compromised OS safe.
• Destinations and external copies: Approved services receive values. Check the recipient’s sharing settings, permissions and policies. External copies such as documents, backups and screenshots, and values already sent, are not recalled by locking the vault or disconnecting AI.
• External information and inference: AI may still learn or infer information from public sources or permitted task results.

Programs and child processes receiving values through environment variables or temporary files can read the actual values and write them to files or send them over the network. This includes AI-written code. Output redaction reduces accidental disclosure of supported value forms; it does not guarantee blocking arbitrary transformations or exfiltration.

When you allow project control, selected files and command results are sent to the AI. Commands run with your PC user's permissions and can access files outside the selected folder and the network. Selecting a folder does not sandbox commands.

Automatic approval skips the prompt for each execution request. AI activity can extend automatic locking; lock the vault manually when you need work to stop while you are away. If a value is saved to an approved service, the AI may be able to retrieve it by reading that service later.

If you suspect a compromise

Stop AI tasks, lock the vault and revoke the connection and other tools’ access. If the device may be compromised, isolate it from the network and revoke or replace service tokens and passwords from a separate trusted device. Information already sent must be addressed separately with the recipient. Do not attach raw secrets, master passwords or vault files to support requests.

Liability and your rights

To the extent permitted by law, tinygem is not liable for losses caused by users, third parties or external conditions where tinygem is not at fault. This does not exclude statutory liability for intentional misconduct, gross negligence, product defects or breaches of security obligations, or users’ statutory rights. For details, see Section 5 of the Terms.