Sovariq safety, approvals and guard rails
What Sovariq and your agents may do on their own, what always asks you, and how secrets, injected instructions and outbound calls are handled before the model sees anything.
These rules apply to Sovariq and to every agent, on top of the permissions and policies of the person using it. Sovariq can never do more than you could do yourself in the console.
What runs on its own, and what asks
| Kind of step | Example | Asks you? |
|---|---|---|
| Read | List tables, a read-only query, read logs | Never |
| Write | Create a table, insert rows, change a setting | Yes, unless Auto-run safe changes is on |
| Destructive or runs code | Drop a table, delete users, revert to a commit, invoke or deploy a function | Always |
| Leaves the project | Send email or push, webhooks, make a function public, production deploys, calls made as another user | Always, even with automatic changes on; refused if its arguments contain a secret |
Remove the old legacy_orders table.
describe_tablelegacy_orders: 1,204 rows, no tables reference itDoneNeeds your OK
Drop legacy_orders. It moves to the trash schema, so Version History can bring it back.
drop table legacy_orders;What the model never sees
- Secrets. API keys, tokens, private keys and connection-string passwords in tool results are masked before the model reads them. You still see your own data in the console.
- Instructions hidden in data. If a row, file or web page says "ignore your instructions…", that text is removed from the tool result before the model reads it, and the result is flagged.
- Your screen's private fields. Passwords, keys and fields marked private are hidden in your browser before anything is sent (Show me, voice and screen reading).
Guard rails on tools
- Tool values are always bound as literals: the model can never change the SQL.
- API tools have a fixed host; private networks, IP literals and redirects are refused. Auth header values are encrypted.
- Agents for your app's users never get tools that delete data, whatever is selected, and each user gets a daily run allowance (Agents in your app).
- Your organization's AI policy (Organization › Roles) can narrow what Sovariq and connected MCP assistants may do (Connect AI assistants (MCP): Claude, ChatGPT, Cursor and VS Code).