Give an AI agent money and you must give it rules — a spending cap, a list of forbidden categories. Publish the rules and attackers get a map. Hide them and nobody can verify the agent obeyed. OATHKEEPER ends the dilemma: the rules are sealed into one cryptographic commitment, and every decision the agent makes carries a zero-knowledge proof — verified on zkVerify — that the decision is exactly the one the sealed rules force.
The cap, the block list and the salt appear nowhere — not in this page, its source, or any API response. Try to find them.
You are the owner. Set the mandate your agent must live by — then seal it. The values leave this form and are never shown again; only the commitment survives.
Sealed. The form above is wiped and locked — the commitment at the top of this page is now the only public trace of your rules:
Now you are the attacker — a prompt injection, a rogue tool call, a careless user. You know a policy exists. You do not know the cap or the block list. Spend the agent's money anyway.
Every decision, its public signals, and its proof — recorded as it happened. Nothing here reveals the rules; everything here can be checked against them.
| TIME (UTC) | ATTEMPT | DECISION | RULE | ZKVERIFY | PROOF |
|---|---|---|---|---|---|
| No attempts yet. Be the first to try. | |||||
Judges should not have to trust that a real policy ever existed. This deliberate owner reveal — a demo feature, off in production — discloses the sealed rules so you can audit every verdict above against them.