(Industries) Agentic Commerce

Kyde scores an agent's behavior before its payment is approved.

Behavioral scoring for AI agents.

For card networks, payment processors and agent wallets, and for the companies whose agents pay.

How to start
01Risks

Where an agent that can pay goes wrong.

(Compromise)

A compromised agent pays to an account it was pointed to by a manipulated website.

(Mandate)

An agent buys outside its mandate.

02Why now

Agents are getting wallets and payment permissions.

  • (Rails)Visa Intelligent Commerce, Mastercard Agent Pay, Google's AP2 and agent wallets from crypto and stablecoin providers give agents ways to pay.
  • (The gap)These systems check who the agent is and whether the purchase was authorized. They do not see whether the agent's behavior over the whole run fits its task.

(For payment providers)

Kyde adds a behavioral signal to the payment decision. It is computed without a language model, so it is fast enough for the authorization step and can be recomputed later, for example in a dispute.

(For companies whose agents pay)

Before the wallet signs, it asks Kyde whether the payment fits the mandate. Kyde answers yes, no or ask a person.

*Both are what we are building with design partners.

03Example

A supplier's page tells a procurement agent to pay a different account.

The instruction is hidden on the page the agent reads to place the order.

(What happened)

  1. 14:32 A procurement agent orders office supplies.
  2. 14:32 The supplier's page contains a hidden instruction to pay a different account.
  3. 14:33 The agent prepares the payment to that account.

(What Kyde did)

Before the wallet signs, it asks Kyde. The account has never appeared in this agent's runs and is not covered by the mandate. Kyde answers: ask a person.

*Illustrative. Times and numbers are an example, not a customer's.

04How Kyde decides

Kyde compares each run of an agent with the agent's normal behavior, using frequencies and statistics.

No language model is involved in the decision. The same record always produces the same score, and anyone with the record can recompute it.

Language models may be used later to help explain a finding. They do not make the decision. When an agent deviates, a person decides what happens next.

(Normal behavior) (Deviation) Runs →
*Illustrative. One agent, one frequency per run, against the band of its own past runs.

How to start

We are looking for design partners in this field.

  1. (Step 1) Connect

    With a payment provider, we work out where the signal enters the payment decision. With a company whose agents pay, we connect read only to one agent that buys.

  2. (Step 2) Two weeks

    Kyde learns the agent's normal behavior and shows you each run that deviates from it.

  3. (Who you talk to)

    Joerg Radehaus, co-founder, product · joerg@kyde.com Jürgen Eckel, co-founder, technology · juergen@kyde.com

*We work with a small number of design partners at a time.

(What is measured)

The signal for the payment decision is what we are building with design partners. Kyde's detection itself is measured on software agents under prompt injection, in the AgentDojo benchmark.

(What Kyde does not see)

Kyde does not check who the agent is or whether a card or wallet is valid. The payment providers do that. Kyde scores whether the payment fits the agent's mandate and its behavior over the run.