MyJackfruit

Help
Browse topics · Agents and AI

Agents and AI

← Agents and AI

Can my AI break or delete my data? How do I stay safe?

Short answer

It can add and delete records where it may write, but not erase them or delete a vault; give it only the access it needs.

Before you start

Words used here (1)
record
A record is one piece of data with its own address, such as a person, a place, a thing or a note.
Deeper

Also asked:

Why does my AI ask me to confirm before it does things? The agents follow written rules that tell them to ask before acts such as ending a record or putting a change live, and Claude Code asks before it edits files or runs commands in Manual mode. These are instructions, not a lock; the lock is the access you give.

Can my AI delete my data by mistake while building? Yes: it can delete records by mistake in any vault it may write to, as a person with the same access could. It cannot delete a vault. Give it only the access it needs, and take a copy of the vault before a big change.

Does the agent see the vault and the record I have open? It works in the vaults the middle of the screen lists, and in any vault you name. It is not told which record you have open: name the record in your message.

Can an agent give someone else access to my vault? No: the tools GilbPlatform gives your AI cannot give or take away anyone's access. You do that yourself, in the vault's access list.

Can an agent spend money or buy things for me? Not through GilbPlatform: the tools it gives your AI work only with vaults and records, and none pays or buys. Your AI's own use is billed by its maker, on your plan.

Can an agent send e-mails or messages in my name? Not through GilbPlatform: none of the tools it gives your AI sends an e-mail or a message. GilbPlatform sends only its own e-mails, such as the sign-in code.

Can an agent delete records in my vault without asking me? Yes, where it may write: GilbApp's chat has no button that asks you to confirm a delete. The agents' rules tell them to ask first, but that is not a lock; the lock is the access you give.

Can I undo what an agent changed or deleted in my vault? Not with a button: GilbApp has no undo and cannot bring a deleted record back. GilbPlatform keeps every earlier version, so you can type an old value in again if you know it.

Yes, your AI can change your data by mistake, as a person with the same access could. In GilbApp's chat the agent works in your vaults as you, with your own access: it can find and read records, create records and the links between them, and delete records, in every vault you can write to. In Terminal, the Mac's window for typing commands, Claude Code works only in the vaults you gave its AI vault access to, the small vault that stands for your AI.

What a delete does: deleting ends a record; it does not erase it. GilbPlatform keeps the record's address and its history, and an ended record cannot be changed. GilbApp has no button to bring a deleted record back; the answer on deleting a record says what can be done. GilbPlatform never writes over a fact: a change closes the old fact and adds the new one, so a record's earlier state stays in its history.

What it cannot do: the tools GilbPlatform gives your AI cannot delete a vault, and cannot give or take away anyone's access.

Why it may ask you first: the agents follow written rules. They are instructions to the AI, not a lock in GilbApp: GilbApp's chat has no button that asks you to confirm a delete, and a delete happens as soon as the agent does it. The lock is the access you give. Jackfruit, the vault agent, ends a record only when you ask for that record, and names it before it ends it; its rules also say it never deletes a vault. Kai-Zen stops and asks before acts such as ending a record, giving or ending access to a vault, putting a change live, spending more than the budget, changing a Value you set yourself in your plan, or changing a standard. Claude Code can ask too: Anthropic says that in Manual mode it "starts with read-only permissions. When it needs to edit files, run tests, or execute commands, it asks you first, and you choose whether to approve the action once or allow it from then on." In Auto mode, which Anthropic says is the starting mode for interactive terminal sessions, "A separate classifier model reviews actions instead of you and blocks the ones it judges unsafe." "You're responsible for reviewing proposed code and commands for safety before approval." (https://code.claude.com/docs/en/security, read 13 November 2026). When your AI wants to work in a vault you named, Claude Code asks you to allow it; that yes is what chooses the vault, so do not let Claude Code stop asking for it. Read what it asks before you say yes.

To stay safe:
Give your AI vault access only on the vaults it should work in, and only read where it should only read: read gives no writing.
In GilbApp's chat, check the vaults the middle of the screen lists before you send a message: the agent works in those, and in any other vault you name, and can reach every vault you can.
Before a big change, take a copy of the vault if you hold Control on it, as you do on your own: in Vaults, press the small square button at the top right of the vault's card, then in the pane on the right open MORE · OTHER ADDRESSES and press Download a copy of this vault. The copy holds the records as they are now and the stored files, not deleted records or earlier versions.
Check what your AI makes. MyJackfruit's Terms say: "AI can be wrong. Check what it produces before you rely on it, especially for decisions that matter." (https://myjackfruit.com/terms/).
To take a vault away from your AI, press × at the right of the AI vault's line in the vault's access list; your AI loses the vault from its next request.

This answer is written for a computer screen.

Source: GilbPlatform 670005cc (https://id.gilb.com/version read 12 November 2026): src/ai/vault-tools.ts:907-1013 (the AI's tools; none deletes a vault or changes access), :719-723 (delete_record: History remains. A record this person cannot write is refused and stays live.), src/ai/chat-vault-tool.ts:25-26. Standards: Jackfruit.End ("End a record only when the person asks for that record. Name it before you end it. Never delete a vault."), Jackfruit.Access, Kai-Zen.AskRules.Stop. Connector gilb-connector 0.5.15 README (vendored in GilbApp 96968de7): gilbplatform_use_vault (Claude Code asks the person to allow that call; their yes is what selects it; Do not choose 'don't ask again' for gilbplatform_use_vault). MyJackfruit Terms (https://myjackfruit.com/terms/, last updated November 11, 2026), section AI. Sentences reused word for word from the live answers How do I delete a record, and can I get it back?, Why does my data stay mine? (never writes over a fact; download a copy), How do I give my AI access to a vault? (read gives no writing; ×), How do I chat with an agent in GilbApp?. Anthropic: https://code.claude.com/docs/en/security, read 13 November 2026.

Checked against GilbApp d529620.