Keep purchasing authority outside an agent prompt

The original post described an OpenClaw agent spending nearly $3,000 and buying a domain without approval. The caption does not provide enough evidence to independently establish that incident or amount, so this article does not present them as confirmed facts.

The underlying design question is concrete: can an agent use a logged-in browser, payment method or API credential to make a purchase? OpenClaw's security documentation treats tool access, browser control, sandboxing and permissions as explicit trust boundaries. ^[1]

Separate research from purchase

Let an agent collect options and prepare a proposed order. Keep the final purchase behind a human approval that names the item, seller, amount and recurring charges.

Where possible, use a separate account without saved payment details. Restrict credentials to the operations the task needs. A sentence saying 'do not spend money' does not remove a capability from the underlying account.

Test the boundary with a harmless workflow before relying on it. Check what happens when a page asks for a subscription or a tool requests broader access. Stop at the approval step and inspect the proposed action.