There is no dedicated setting for this yet. What you can do is limit what the agent can see: connect a development workspace instead of production, sample from tables that hold no sensitive values, and keep the committed test fixtures synthetic.
The agent needs schemas to write correct code, so table and column names are usually part of a prompt. Any controls we add are documented on the security page. Until then, assume a prompt can contain anything in the tables you point the agent at.
See what the model provider sees and the data handling docs.