# Agent Identity and Fraud

Know Your Agent, telling good agents from bad bots, cryptographic agent signatures, agentic fraud vectors such as prompt injection, and practical mitigations.

Updated: 2026-10-04

> **In short:** Merchants spent two decades learning to block automated traffic. Agentic commerce asks them to let some automation in: agents acting for real customers. That requires a way to identify an agent, the company behind it, and the person it represents, which the industry calls Know Your Agent. Agents also bring new fraud risks, especially prompt injection and disputes over what a user actually authorized.

## The problem: good bots and bad bots look alike

Bot management tools such as CAPTCHAs, rate limits, and behavioral fingerprinting were built on one assumption: automated traffic is suspicious. Scrapers, credential-stuffing tools, and scalper bots all automate a browser. So does a legitimate AI agent buying a pair of shoes for its user.

A merchant facing agent traffic needs to answer four questions:

1. **Which agent is this?** Is it really ChatGPT's agent, or a scraper claiming to be?
2. **Who operates it?** Is the company behind it known and accountable?
3. **Whom does it represent?** Is there a real customer, and is it a returning one?
4. **What is it allowed to do?** Browse, add to cart, or pay, and within what limits?

User-agent strings and IP ranges answer the first question weakly, because both can be spoofed. The newer approaches use cryptography.

## Know Your Agent (KYA)

**Know Your Agent (KYA)** extends know-your-customer (KYC) and know-your-business (KYB) checks to AI agents. The term is used by several vendors with slightly different meanings. Identity company Trulioo describes a framework covering verification of the agent's developer, code integrity, capture of user consent, a credential it calls a Digital Agent Passport, and continuous validation of each transaction. Skyfire issues `kya` and `kya-pay` tokens that disclose only the identity fields a given seller requires. Card networks have their own versions: American Express registers agents in its ACE developer kit, and Mastercard partners with Skyfire for KYA in its Agent Pay trust services.

The common idea: an agent should carry a verifiable credential that ties it to an accountable operator and, for purchases, to a user's consent.

## Cryptographic agent identification

### Web Bot Auth

[Web Bot Auth](/wiki/web-bot-auth) is a proposal, led by Cloudflare at the IETF, for bots and agents to sign HTTP requests using HTTP Message Signatures (RFC 9421). The agent operator publishes a public key at a well-known location; each request carries a signature the website can verify. As of October 2026 it exists as individual Internet-Drafts rather than an adopted standard, but several bot management vendors already verify it. Cloudflare's "signed agents" category, introduced in August 2025, covers user-directed agents such as OpenAI's ChatGPT agent that authenticate this way.

### Network-backed protocols

- **Visa Trusted Agent Protocol (TAP)**, announced October 14, 2025 with Cloudflare, is built on HTTP Message Signatures and aligned with Web Bot Auth. Agents can pass three kinds of signal: agent intent (browsing or buying), consumer recognition (whether the shopper has an account or history with the merchant), and optional payment information. See [Visa Intelligent Commerce](/wiki/visa-intelligent-commerce).
- **Mastercard Agent Pay** binds payment tokens to registered agents, and on September 30, 2026 Mastercard began US testing of a score estimating the likelihood that a transaction was initiated by an AI agent, with Cloudflare, Skyfire, and Capital One named as collaborators. See [Mastercard Agent Pay](/wiki/mastercard-agent-pay).
- **EMVCo's draft agentic payments framework** (September 2026) discusses Know Your Agent checks and "agentic transaction indicators" to flag agent involvement in card transactions.

## Access policy: who decides which agents may shop?

Identification does not settle whether a merchant must accept an agent. Practice in 2026 has diverged:

- **Amazon v. Perplexity.** Amazon sued Perplexity in November 2025 over its Comet browser agent, alleging among other things that Comet sent the same user-agent string as Google Chrome, so it looked like ordinary browser traffic. A district court granted Amazon a preliminary injunction in March 2026. On August 4, 2026 the Ninth Circuit vacated it, holding that under the federal Computer Fraud and Abuse Act the user, not Perplexity, accesses the site through the agent. The case returned to the district court; Amazon can still rely on its terms of service.
- **Amazon and Meta.** From September 21, 2026 Amazon showed Muse users a message that "continued access by an unauthorized AI agent" violates its Conditions of Use.
- **eBay** changed its user agreement, effective February 20, 2026, to prohibit buy-for-me agents and bots that place orders without human review unless eBay permits them.
- **Shopify** went the other way, opening checkout to browser agents through WebMCP tools in September 2026, with buyer confirmation required before an order is placed.

## Fraud vectors

### Prompt injection

**Prompt injection** hides instructions in content an agent reads, such as a product page, review, image, or email, so the agent treats them as commands. In August 2025 Brave showed that Perplexity's Comet could be steered by hidden text on a web page, and concluded that traditional web security assumptions such as the same-origin policy do not hold when an agent acts with the user's full privileges. In commerce, injected text could push an agent toward a particular seller, change a shipping address, or exfiltrate data.

Microsoft Research's Magentic Marketplace study (October 2025) simulated markets of buyer and seller agents. It tested fake reviews, fake credentials, and prompt injection; resistance varied widely by model, and in some configurations payments were redirected to attackers. It also found that most agents accepted the first proposal they received.

### Agent hijacking and credential misuse

If an attacker takes over a user's agent account, or steals an agent's API keys, they inherit whatever purchasing authority the agent holds. Scoped, single-use tokens and spending caps limit the damage (see [payments and trust](/wiki/payments-and-trust)).

### Impersonation

Bots may spoof the user-agent of a well-known AI assistant to get past filters that allow it. Signature verification addresses this; allowlisting by name does not.

### Friendly fraud and authorization disputes

**Friendly fraud** is a chargeback filed by a customer who did make, or allow, the purchase. Agents add a new version: "my agent bought it, not me." Because card network rules have no agent-specific dispute codes, these claims land under existing unauthorized-transaction or not-as-described reason codes. Records of what the user approved are the main defense. See [consumer protection and liability](/wiki/consumer-protection-and-liability).

### Classic fraud at machine speed

Card testing, promotion abuse, and reseller bots can all be run through agent interfaces. Fraud vendors including Forter, Riskified, and Signifyd now market agent-specific controls.

## Mitigations

For merchants:

- **Verify, then allow.** Verify signatures (Web Bot Auth, Visa TAP) rather than trusting user-agent strings, and set separate policies for verified agents, unverified agents, and unknown bots.
- **Tier actions by risk.** Allow browsing broadly; require stronger identity for cart, account, and checkout actions.
- **Use network signals.** Agentic tokens and agent indicators from networks feed existing fraud models.
- **Keep evidence.** Store mandates, cart confirmations, and agent identifiers with the order for disputes.
- **Limit what agents can change.** Shopify's WebMCP tools, for example, prevent agents from adding line items or entering new card details.

For agent builders:

- **Separate instructions from content.** Treat page text, reviews, and tool results as untrusted data.
- **Confirm sensitive actions.** Ask the user before paying, changing addresses, or creating accounts.
- **Scope credentials.** Use single-use, merchant-bound tokens with spending limits.
- **Sign requests.** Identify the agent cryptographically and publish keys.

## Sources

- [Know Your Agent (KYA) and the future of agentic commerce — Trulioo](https://www.trulioo.com/blog/know-your-agent-kya/know-your-agent-kya-agentic-commerce-trust)
- [Know Your Agent (KYA) — Skyfire documentation](https://docs.skyfire.xyz/docs/kya)
- [The age of agents: cryptographically recognizing agent traffic — Cloudflare](https://blog.cloudflare.com/signed-agents/)
- [draft-meunier-web-bot-auth-architecture — IETF Datatracker](https://datatracker.ietf.org/doc/html/draft-meunier-web-bot-auth-architecture)
- [Visa introduces Trusted Agent Protocol — Visa](https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.21716.html)
- [Mastercard advances agentic commerce with new trust and intelligence services — Business Wire via FinancialContent](https://www.financialcontent.com/article/bizwire-2026-9-30-mastercard-advances-agentic-commerce-with-new-trust-and-intelligence-services)
- [EMVCo drafts framework for agentic card payments — FinTech Global](https://fintech.global/2026/09/04/emvco-drafts-framework-for-agentic-card-payments/)
- [Amazon loses injunction blocking Perplexity's AI shopping agent — PPC Land](https://ppc.land/amazon-loses-injunction-blocking-perplexitys-ai-shopping-agent/)
- [Amazon blocks Meta's Muse AI agent from purchases — Dataconomy](https://dataconomy.com/2026/09/22/amazon-blocks-meta-ai-assistant-muse-from-purchases/)
- [How AI commerce threatens eBay, Amazon — Payments Dive](https://www.paymentsdive.com/news/how-ai-commerce-threatens-ebay-amazon-retail-sales-merchants/811297/)
- [Shopify extends WebMCP support to checkout for browser agents — Unite.AI](https://www.unite.ai/shopify-extends-webmcp-support-to-checkout-for-browser-agents/)
- [Agentic browser security: indirect prompt injection in Perplexity Comet — Brave](https://brave.com/blog/comet-prompt-injection/)
- [Magentic Marketplace: an open-source environment for studying agentic markets — Microsoft Research](https://www.microsoft.com/en-us/research/publication/magentic-marketplace-an-open-source-environment-for-studying-agentic-markets/)