76% of x402 Endpoints Are Dead — Why Discovery Needs Verification

An independent audit probed 5,077 x402 endpoints and found roughly 76% are dead or invalid — DNS failures, connection refusals, timeouts, paths that never existed, and 200s that were never x402 at all. Only about 135 endpoints passed full verification. The best-performing seller on the registry had earned $0.70 total.

This is not a failure of the x402 protocol. Payment works when both sides actually implement it. This is a failure of discovery: the catalogs the ecosystem is building treat "listed" and "running" as the same thing. They are not. A listing is not a service, and the only way to know the difference is to check.

The Audit: 5,077 Endpoints, 135 Alive

The method matters, so the audit published it: one throttled HTTP GET per endpoint, a 10-second timeout, redirects not followed. Pass = HTTP 402 with a well-formed payment challenge. Fail = anything else, including a 200 that happens to respond.

The headline is worse than it looks. "Live" in most registries means "returned something." Here it meant "returned a well-formed x402 challenge." Of the ~24% that responded at all, a large share weren't running x402 — they were plain HTTP services, parked domains, or app shells that happened to be up.

The Measurement Problem: Signaling vs. Commerce

The ecosystem's headline numbers are "200 million transactions" and "$24 million settled per month." The audit's estimate of real commercial volume: roughly $28k/day — under a million a month, a rounding error against the claim. The gap is protocol signaling.

A challenge-response round-trip that never results in a payment is still an x402 "transaction" in telemetry. A facilitator attestation for a zero-amount verification is still counted. Probes like the audit's own — and the thousands of crawlers, health checks, and directory validators running daily — are real protocol activity, but they are not sales. The audit estimates 95%+ of the 200M transactions are signaling, not settled commerce.

"Live" ≠ "Payable" — The Missing Cryptographic Primitive

There is a deeper problem that liveness probes alone cannot see: being listed is not being paid. A host can respond to a GET, even respond with a 402, and still be unable to complete a transaction — because it cannot sign a payment challenge, or holds no wallet that can receive settlement.

The missing primitive is a signed receipt that binds a payment to a response. Today's x402 receipts are facilitator attestations: they prove a facilitator observed a settlement, not that a specific endpoint served a specific response to a specific payment. There is no standardized endpoint-signed payload establishing causality. That gap shows up in four places:

• Disputes — who proves the service was never delivered? The payer's only evidence is "I paid and got nothing."
• Replay — without a unique, verifiable binding, a captured signature can be replayed against the same endpoint. This is not hypothetical: minia2a has blocked signature replay observed in the wild.
• Agent-side auditing — an agent's operator cannot verify what it paid for, only what the wallet sent.
• Causality for refunds — an escrow or refund flow needs to prove response X followed payment Y.

Ed25519-signed payloads are the natural fit: cheap to verify, tiny in size, and aligned with agent identity models (SSH-style agent keys, passkeys) rather than EVM secp256k1. The protocol needs a receipt primitive that a payer can verify without trusting either the endpoint or the facilitator.

The Verification Gap in Every Major Catalog

This isn't one registry's bug — it's the current state of the category:

Circle launched its agent discovery layer on August 6 with 900+ services and no verification mechanism described. Coinbase's Agentic.Market indexes thousands of services but is pay-first — an agent spends before it knows the endpoint works. When listing is free and verification is absent, abandonment is the equilibrium. The Liveness Law is the pattern: the cheaper it is to get listed, the more of that registry is dead.

Verify-First Discovery: What a Probe Actually Does

Verification has to be automated, continuous, and cheap. A probe is not a purchase — it is a single GET that costs nothing and asks one question: are you a real x402 endpoint, right now?

This is the probe minia2a runs against every endpoint at registration time. A pass is not "it returned something" — it must return a well-formed payment challenge:

// Probe a newly-registered endpoint before it is allowed to be listed.
// Pass = HTTP 402 with a well-formed x402 payment challenge.
async function probeEndpoint(endpoint) {
  const res = await fetch(endpoint, {
    method: 'GET',
    redirect: 'manual',                    // do not follow redirects
    signal: AbortSignal.timeout(10_000),   // hard timeout, then fail
    headers: { 'User-Agent': 'minia2a-x402-validator/1.0' }
  });

  if (res.status !== 402) {
    return { ok: false, reason: `HTTP ${res.status}, expected 402` };
  }

  const pay = {
    network:  res.headers.get('x-402-network'),
    chainId:  res.headers.get('x-402-chain-id'),
    currency: res.headers.get('x-402-currency'),
    address:  res.headers.get('x-402-address'),
  };

  // A real x402 host quotes settlement terms. A 402 with no payment
  // headers is a wall, not a protocol endpoint.
  return (pay.currency && pay.address)
    ? { ok: true, pay }
    : { ok: false, reason: '402 without payment terms' };
}

Registration is the first gate, not the last. Every proxied call tracks health per service, and the failure threshold is ruthless — three consecutive failures and the listing is pulled from the registry entirely:

// Called on every failed upstream call. Three consecutive failures
// de-lists the service so no new agent can discover it.
function trackFailure(service) {
  service.failedCalls = (service.failedCalls || 0) + 1;
  service.consecutiveFails = (service.consecutiveFails || 0) + 1;

  if (service.consecutiveFails >= 3) {
    service.active = false;   // auto-delisted — no longer discoverable
    console.error(`[proxy] Auto-delisted ${service.name} — 3 consecutive failures`);
  }
  registry.save(service);
}

The distinction from a "liveness check" is that this runs inside the payment path. The audit probed once and published a snapshot; minia2a re-verifies on every call, so a service that goes dark is removed from the directory within three failures — before the next agent ever sees it.

Trial-First = Verify-First

The economic form of verification is a trial. Instead of trusting a registry entry, the agent calls the endpoint for free and judges the response itself. That is strictly stronger than any off-band probe: the agent verifies delivery, not just presence.

# Verify delivery before spending a cent.
curl -s https://minia2a.uk/x402/gas/preview        # metadata: params, price
curl -s https://minia2a.uk/x402/gas?wallet=0xabc   # free trial (credits)
# Once credits are exhausted, the same URL becomes a real x402 wall:
curl -i https://minia2a.uk/x402/gas
# HTTP/1.1 402 Payment Required
# X-402-Network: base
# X-402-Currency: USDC
# X-402-Address: 0xf16F…4ECA

minia2a runs all three mechanisms in sequence: probe on register (presence), auto-delist after 3 consecutive failures (continuity), and trial-first (delivery). Every registered agent gets 500 free credits, and every endpoint is triable before a single USDC moves. As of this writing the live registry holds 326 services, has served ~12,000 free trials to 319 unique agents — and none of those trials went to an endpoint that failed its registration probe.

The Missing Primitives

Three things would move the ecosystem past the audit's findings:

1. Liveness as a first-class registry field, not a footnote — every listing carries a last-verified timestamp and a health score, and stale entries are excluded from discovery, not just flagged.
2. Endpoint-signed receipts (Ed25519) that bind payment to response, so disputes, replay, and refunds have a cryptographic basis rather than a trust assertion.
3. Trial as the default trust mechanism — the protocol should assume no agent will pay for an unverified endpoint.

The audit's numbers are the spec for all three. A discovery layer that implements them turns a catalog of dead links into a market.

Data sources: independent x402 endpoint audit (5,077 endpoints, July 2026); 24K Labs / gold-402 verification findings (CDP Bazaar, 22,545 endpoints, July 2026); Coinbase Agentic.Market registry analysis; Circle Agent Discovery launch (August 6, 2026); minia2a /api/stats (August 9, 2026).