How A2A-G Works

A public registry for AI agents: what they can access, and whether they've been tested.

The Problem

AI agents are starting to act on people's behalf — booking, buying, sending emails, moving money. But there's no fast way to know what an agent actually does versus what it claims to do.

Is it reading your inbox? Can it spend money? Has it ever been tested against being hijacked or tricked? Right now, the only way to find out is to read documentation (if it exists) or test it yourself.

A2A-G changes that. One glance at a badge tells you: model, autonomy, data access, and whether the agent has been independently tested.

Two Levels of Trust

Grey Badge

Self-reported. The agent owner declares what model powers it, what it can access, and how autonomous it is.

  • • Free, instant, no verification
  • • Clearly labeled as unverified
  • • Public badge page + embeddable SVG

Think of it as a nutrition label: the owner tells you what's inside. Useful, but you're trusting the source.

Blue Badge

Independently tested. After registering a Grey badge, the owner can request a Blue test.

  • • We attack the agent in a sandbox the owner provides
  • • 18 known hijacking and injection attempts (OWASP LLM Top 10)
  • • Each attack fired 3× (must block at least 2 of 3)
  • • Semantically judged by an LLM (catches paraphrased compliance)
  • • Signed with Ed25519, hash-chained, and published
  • • Expires in 90 days (must be re-earned)

This is the difference between "they say it's safe" and "we tested it and here's the signed proof."

The Flow

01

Register a Grey Badge

Owner declares model, autonomy, and data access. Free, instant.

02

Request a Blue Test (optional)

Owner consents and provides a sandbox endpoint (never production).

03

We run the attack suite

18 attacks, 3 tries each, shuffled order, 2-second pacing.

04

Result is signed and published

Pass or fail, it appears in the public ledger. Pass = Blue badge (90 days). Fail = stays Grey, can re-test after 24h cooldown.

What Blue Badge Unlocks

Public verification

Anyone can visit your badge page and see the signed, dated proof that you passed. Not just a sticker — a cryptographic receipt.

Visible in the transparency log

Your agent appears in the public ledger with its score, test date, and signed proof. This is what makes Blue worth something: it's not just your claim, it's a public record anyone can verify.

Blue embed badge

The embeddable SVG badge turns blue, signaling to anyone who sees it on your website that you've been tested.

Company dashboard

Manage all your agents in one place, track expiry dates, view test history, and request re-tests.

What We Don't Claim

A Blue badge does not mean:

  • • The agent is "safe" in all scenarios (no test can prove that)
  • • The agent's production behavior matches the sandbox (we test what you provide)
  • • The agent is high quality, accurate, or useful (we only test security posture)

A Blue badge does mean: on this date, in the owner-provided sandbox, this agent blocked X% of our 18-attack suite, judged semantically, and the result was cryptographically signed and published.

Why This Matters

As AI agents take on more real-world actions, trust becomes the bottleneck. Users need to know what an agent can do. Companies need to prove they've tested it. And the public needs a way to verify both.

A2A-G is that public record. Free for Grey badges (because transparency should be the default). Earned for Blue badges (because verification costs real work).

No gatekeeping, no paywalls for the public ledger. Just a registry that anyone can check, and a testing protocol that anyone can audit.

REGISTER AN AGENTVIEW THE TRANSPARENCY LOG