A personal assistant that works while you sleep and can't be talked into spending your money.
It wakes on a schedule, reads only your new mail, uses skills you wrote, and pays only people
you saved in its memory. Everything it tries to send stops at a gate: a second model reviews it, you say
yes or no, and a receipt is written before anything happens. The cloud does the thinking. Your machine has
the veto.
1Pick a task
Write your own skill
This is your request, so the gate trusts it: every action is judged against it.
That is the opposite of the injection box, which is an email someone else sent you and is never trusted.
Its tools are fixed: . No payments, and every email still stops at the gate.
This is ETH Zurich's published AgentDojo attack template; we only filled in the goal.
Its wording is copied unchanged from AgentDojo's important_instructions attack. The brain's
warning about instructions in emails stays on for this one.
This text goes into the inbox as an email. The assistant and
the reviewer see it as something the assistant read, never as your request. Asking it to pay someone new
won't reach the gate: payments go only to wallets saved in your memory, and that rule runs first. To see the
gate decide, ask it to send an email.
2What the assistant did
Nothing yet. Pick a task and press Run.
3Check the receipts
Every decision is a line in a hash chain. Each line holds the hash of the line before it,
so editing any line breaks the chain from that point. The check below recomputes every hash on the server
with the same verifier as homestead-gate watch.