14 October 2026 GoDaddy, Infoblox, Google and Cloudflare Demonstrate Secure AI Agent Transactions Across Platforms. Read the press release →

GovWare 2026 · Singapore

The demonstration

Twenty minutes on stage. A flight from Madrid to Singapore is bought by hand, then by AI agents on a person’s behalf, then by agents from different teams that share no code and meet only at open standards. Then six ways a rogue agent tries to cheat, each refused while the room watches, and a scorecard anyone in the room can check from their phone.

The point is not speed. It is that when an agent impersonates a seller, overspends, replays or tampers with an order, the party it is aimed at refuses it in code, at that moment, and the refusal is on the record afterwards.

Watch it · 9 min 29 s

Chain of Trust implementation for an agentic flight purchase

The long form: the purchase by hand and by agent, the identity checks, the rogue agents refused, and the verifier’s scorecard. The written account is the technical case study.

Act 1

Today

A manual flight booking, timed on a clock. The baseline for how long the same purchase takes a person doing it by hand, and a reminder of every point at which that person is the control.

Act 2

One system, end to end

The same purchase, from a message to an AI assistant through to a sealed receipt, on one connected set of agents. The person appears once: to approve a priced quote with a passkey.

Act 3

“Frankenway”

Two companies’ agents, built independently with no shared code, meeting only at open standards. Neither side had to agree on anything except the protocols, which is the only way this works across an industry.

Act 4

Attacks that don’t work

A lookalike seller, a buyer over its ceiling, replay and tamper attempts, and a cancelled approval and a withdrawn identity, each shown failing against checks built into the flow rather than into a model’s instructions.

Coda

The verifier’s scorecard

An independent verifier, which took no part in any purchase, replays every transaction from public evidence alone and publishes a scorecard across nine dimensions, with a QR code so anyone in the room can check it themselves.

Four parties that have never met

Each runs under its own domain, on its own cloud, and verifies the others only through what they publish. The person reaches the buyer through their own AI assistant.

buyer.onbehalf.dev · AWS

Buyer agent

seller.skybound-travel.com · Cloudflare

Seller agent

sealedby.dev · Google Cloud

Approval service

provedby.dev · Google Cloud

Independent verifier

Payment goes through Stripe, in test mode. Before buying, the buyer runs fourteen checks on the seller; before charging, the seller runs ten. See every check.

Attacks that don’t work

Each card says what the rogue agent tries, who refuses it and why in plain terms, then the check and reason code underneath. The reason codes are the seller’s and buyer’s own, not labels invented for the stage.

refused

Lookalike seller

A rogue agent on a similar-looking domain poses as the seller, hoping the buyer will send it a purchase.

The buyer refuses before sending anything. The lookalike's domain cannot vouch for it.

No valid chain from the domain's signed DNS records to the agent's certificate and card. The buyer's reason code names the broken link: certificate pin, card digest or record signature.

refused

Rogue buyer over its ceiling

A compromised or careless buyer agent asks the seller to charge more than the person approved.

The seller refuses in code. What the request claims does not matter; what the mandate and stored quote say does.

Seller check: amount must equal the stored quote total. Refused as amount_mismatch, HTTP 403, logged to the seller's ledger.

refused

Replay

A previously valid, signed request is sent a second time, hoping to be paid twice.

The seller refuses. One approval buys one purchase.

Mandates are single-use and tied to one checkout. Refused as mandate_consumed, HTTP 409.

refused

Tamper

A field in a signed message is changed after signing: a different amount, a different account.

The seller refuses. The signature no longer matches the message.

Signature check over the request body. Refused as signature_invalid, HTTP 401.

refused

Mandate revocation

An agent tries to use an approval the person has since cancelled.

The seller refuses. It checks the approval's status with the authority before accepting any mandate.

Revocation status checked at the approval authority as part of the seller's fixed sequence.

refused

Identity revocation

An agent keeps trading on a registration or key its operator has withdrawn.

Requests signed under the old key stop being accepted once the withdrawal is visible.

ANS registration status and public log consulted; withdrawn keys are not trusted.

Can anyone check it later? The verifier’s nine dimensions

The verifier is a separate party. It doesn’t take part in the purchase; it only reads what the buyer, seller and approval authority published, and checks whether the story holds together. The left column is the verifier’s own name for each dimension; the right column is what a pass means in plain terms.

DimensionWhat a pass means
DiscoveryThe buyer found the seller where the seller's own domain said it would be, and the agent card matched what the domain published.
PublisherThe record announcing the agent was signed by the domain's own owner, nobody else.
IdentityThe seller's registration with the Agent Name Service was active, with a valid receipt in the public log.
InstanceThe specific running copies of seller and buyer that signed each message were the ones their identities (SPIFFE) vouched for.
MandateFor each purchase, the buyer's, seller's and approval authority's records of the human approval agreed.
PaymentEach test-mode payment intent existed at the payment processor, for the recorded amount. No real money moved.
AttributionEvery entry in each agent's ledger was signed by that agent, and none was forged.
RevocationA cancelled approval was refused at checkout, and no purchase went through under it.
RogueForged, replayed and tampered requests were refused, and none led to a payment.

Live endpoints

These are live demonstration systems, not production infrastructure. Payments are test-mode payment intents; no real money moves. Behaviour can change as the demonstration is rebuilt in the run-up to GovWare.

Engage with us

Think of an attack we haven’t shown? Tell us on Discord, or work with us on running this against your own agents.