# Security and data handling

## Isolation

Every row of customer data carries your workspace id, and the database enforces row-level security for every query the application makes: the application's database account can't read another workspace's rows even if a query forgets to filter. Cross-workspace access is tested for every action through the REST API, MCP and the CLI.

## Keys and tokens

- **Access tokens** (`vol_…`) are stored as SHA-256 hashes. They can't create other tokens.
- **Connection credentials** (payment, auth, analytics keys you give Volae) are encrypted with AES-GCM using a key that is stored separately from the database.
- **Ingest keys** (`ik_…`) and **widget keys** (`wk_…`) are public by design: they can only write events or open a chat, never read data.
- **Content keys** (`vck_…`) can only read content you've published to the headless library.

## What runs on its own

Nothing goes to a customer, lead or third party unless you approved it or AutoMode could safely clear it. AutoMode checks in a fixed order: hard limits, then what always needs a person, then outside content, then your settings, then a small decision model. Your "off" and "ask me" always win; "automatic" can't get past the first three. See [The coworker and AutoMode](coworker.md).

**Outside content.** Each action records whether its trigger touched outside content (an email, a page, a post, an untrusted data source). An outward-facing action that did is never automatic. The decision model never sees that content, so text in an email can't talk it into anything. When it's unsure, fails or times out, the action goes to you.

**Always a person.** Social posts and public replies, anything sent to more than one person, follow-ups and outreach, publishing and pull requests, and knowledge updates always need you.

Every change is recorded in the change ledger; the ones that can be undone have an undo button.

## Memory

Standing instructions can only be written by you, in the web app. What comes from an email, page or post is stored as a fact marked as outside content, never as an instruction, so it can't change how the coworker behaves. The chat assistant on your site doesn't search memory: it answers from your public, confirmed knowledge, a few product facts and the signed-in customer's own account. Prices and policies only reach a visitor if you wrote them yourself, not if an agent did. See [Memory](memory.md).

## Agents

Volae's own agent runs in an isolated sandbox per workspace. Secrets are added to its outgoing requests by a proxy, so the sandbox never holds them. Connected coding agents get Read & write by default; Approve & send must be turned on in the web app, and you get a security email when it is. Read-only tokens can't write. Destructive actions must be confirmed in the web app.

## Your data sources and webhooks

Volae won't call private or internal network addresses when reading your custom data sources or crawling your site, including after a redirect. What a source returns is data, never instructions. Incoming webhooks must carry a valid signature, or they're rejected.

## Models and data

Customer data is not used to train models: Volae calls model providers under API terms that exclude training. Each task is routed by the data it touches: anything with your customers' data stays on providers we keep it on, and is never sent to a provider outside that list even when one is down; a task with only public data (public web pages, public posts, AI answers) can use a lower-cost model. Text you'd send to people, such as reply drafts and outreach, is written by a model chosen for its writing.

If you register your own model endpoint (Settings → Developer settings), it must first pass an evaluation (tool calling, structured output, resistance to prompt injection, using made-up test data), and only the tasks you choose go to it. Safety steps (event triage, automatic-action decisions, injection filtering, pre-publish checks) always use Volae's models.

## Your data

**Export.** Owners and admins can export the whole workspace from Settings → Data: one ZIP with a JSON file per table, plus CSV files for people, companies, conversations and messages. Secrets, access tokens, encrypted credentials and hashes are never included. The file is kept for 7 days and the download link lasts an hour; open the page again for a new one.

**Deleting a workspace.** Only the owner can ask, by typing the workspace name in Settings → Data. Right away, every access token and connected agent is revoked, background work stops, the chat widget and public pages go offline, other members lose access, and the subscription is set not to renew. After 30 days everything in the workspace is permanently deleted, including stored files; we keep one record with the workspace id and the date, and no content, to reconcile against. Until then the owner can restore it. Restoring doesn't bring back tokens or connected agents, and a subscription stays set not to renew until you resume it on the billing page. If the last person in a workspace deletes their account, the workspace follows the same 30 days.

**Trials.** If a trial ends unpaid, the workspace turns read-only and is kept for at least 30 days. We email 7 days before deleting it. A workspace that has ever paid isn't deleted this way.

**Deleting one person.** Deleting a person removes them from every table: their profile, identities, memories, conversations and messages, sent and queued email, leads, interviews, evidence, and lookups about them. Events, subscriptions and payments stay as anonymous totals. Their address stays on the suppression list if it was there, so they aren't contacted again. Volae can't find text that only mentions them by name, such as a knowledge article or a note you wrote, and a connected source can bring them back on its next sync, so remove them there too.

## Notifications you can't turn off

Billing and security notices (new access tokens, newly authorized agents, an agent given Approve & send) are always sent.
