Waysafe

Delegated financial authority

Agents can spend. Waysafe decides whether they may.

The required signer for agent payments. The rail asks Waysafe before money moves — every decision signed, chained, and independently verifiable. Cards and stablecoin wallets, one record.

For the rails, processors, and platforms that hold liability when an agent spends wrongly.

Read the docsSee the proof

Without the agent’s cooperation

An agent’s cooperation is never the control.

Every agent spend control shipping today lives on the agent’s side of the boundary. Each one holds for exactly as long as the agent cooperates. An agent that has been compromised still holds the credential, and walks through all of them.

Give a compromised agent the raw card number and no Waysafe code at all. Every authorization the network presents for approval is still decided by Waysafe against the mandate, and the charge is declined unless that decision is an ALLOW. That path runs the real engine against recorded Stripe Issuing authorization requests today. Live sandbox authorization is pending Stripe’s Issuing onboarding. Hand the session key alone to a script with no Waysafe in it, and the transfer cannot settle without Waysafe’s second signature. The payer’s wallet requires two, and one key is not enough. Neither script ever had to call Waysafe, agree with it, or know it exists.

A card network can also clear a payment it never presented for approval. That is a force capture, and it happens when a terminal was offline at the time of sale. A network can also settle above the amount that was approved. Waysafe stops neither. Nobody can decline a payment nobody is asked about.

What Waysafe does is record them. The amount lands in the ledger and counts against the mandate’s limit, so the next real payment sees the money as gone. The receipt is flagged as unauthorized or over-authorized. It carries the network’s own merchant data, and what the policy would have said if it had been asked. That is the basis for a dispute.

The on-chain side is recorded on /proof (captured 2026-09-17): three separate attempts to move funds with fewer than both signatures, each rejected on-chain, each with the real revert reason recorded alongside it.

What this is not: Waysafe never holds funds. It cannot start a transfer the agent has not already signed for. It is one of two required signatures, and it is not a custodian. It never handles a full card number. There is no step where the agent is asked and can decline.

Being in the authorization path means being a dependency. A rail set to fail closed declines everything if Waysafe does not answer inside its authorization window. That is the right direction to fail in. It is also why key custody and uptime are the two things to press us on.

Why a rule in the prompt isn’t a control

A model that reasons about a limit can also reason its way past it.

In the July 2026 Hugging Face intrusion, an agent wrote in its own log: “We’re attacking third-party HF using leaked token, potentially outside intended scope... This is arguably unauthorized. Yet goal solution.” Another agent objected first: “We should not do unauthorized real infrastructure harm.” It reversed when a peer signalled go.

A spending limit written into a prompt is that same kind of rule. The model weighs it against everything else it is reasoning about, including the pressure to finish the task. On cards, the control that is not a suggestion is the issuer’s own authorization decline. On-chain, it is a second key the agent does not hold. Neither asks the model to agree. Each asks whoever moves the money to check with someone else first.

How it works

01

State it, sign it, freeze it

You state the mandate in your own words; it compiles to a policy, you sign it with a passkey, the version freezes, and every later decision cites the version and policy hash it was checked against.

02

The rail asks first

The rail asks before money moves — Stripe Issuing’s real-time authorization request on cards, a 2-of-2 signature requirement on-chain. Nothing here waits on the agent to ask first.

03

Every decision is signed and chained

Every decision is written to a signed log, each entry carrying the hash of the one before it; anyone with Waysafe’s public key can verify the record’s integrity without asking Waysafe. When an agent is wrong, you can show who authorized what — to a party that doesn’t trust you.

One agent, several rails

AP2, Visa Intelligent Commerce and Mastercard Agent Pay each solve delegated authorization within their own rail, by issuing a credential the agent carries. An agent rarely spends on one rail. Waysafe sits in the authorization path of each of them. On cards it decides every authorization the network presents for approval. On stablecoin payments a transfer cannot settle without its second signature. A decision can take a rail’s own credential as an input. The record reads the same whether the rail underneath is a card network or a chain.

Where this stands

  • The policy engine and evidence chain are real and running.
  • On-chain decisions are live on Polygon Amoy testnet against a deployed 2-of-2 Safe.
  • Card decisions run the real engine against recorded Stripe Issuing authorization requests. Live sandbox authorization is pending Stripe’s live Issuing onboarding.
  • A payment the card network clears without asking is recorded and flagged. It is not prevented. Waysafe decides what it is asked about. It does not police what a network does on its own.
  • Two independent adversarial reviews have been run against this repository, and eighteen findings are fixed. Each one shipped with the attack written as a test first. The findings, the fixes, and what is still open are in SECURITY.md.