did:atp and AAuth Standards Launch — How to Safely Delegate Authority to AI Agents

We are rapidly approaching a world where AI agents buy stocks and manage emails on our behalf. However, handing over our precious passwords or API keys entirely to AI is far too risky. To ensure AI can act as our safe representative, internet platforms have begun to restructure their identity verification systems and security standards.

From Chatbots to Dedicated Members — Why Platforms are Creating 'Agent Accounts'

Until now, for an AI agent to buy stocks or organize emails for us, we had to rely on unstable methods. This was because the structure involved risks like borrowing a user’s entire login session or API key. It was essentially like handing over your front door key and bank security card just to ask someone to clean your house.

To solve this, platforms have started issuing independent, agent-specific accounts directly to AI. A prime example is the financial platform Robinhood, which introduced AI agent accounts that can programmatically control spending limits. By clearly defining the boundaries of funds the agent can handle and providing the ability to block the agent at the API level if abuse is suspected, they have established robust safety measures.

Nylas, an enterprise infrastructure platform, has also introduced an agent account system with enhanced isolation capabilities. Instead of an agent sharing a human’s inbox and peeking at important information, it is assigned a dedicated, isolated inbox and calendar. This allows enterprises to mitigate data leakage concerns while transparently logging and tracking every action performed by the agent.

AI is no longer just a simple assistant tool; it is being treated as a full-fledged participant with clearly defined authority boundaries. With these safe guardrails in place, a practical foundation is being built where AI agents can confidently handle more complex and important tasks.

did:atp and AAuth — Technologies for Safe Authorization in Background Environments Without Human Intervention

Think about the common Google login window or email verification screen we see when making payments or signing up for services online. These methods require the user to manually click a button and enter a verification code. But what if an AI agent needs to judge and book a train ticket or clear out your emails while you sleep? The moment a login or identity verification screen appears, the agent stops working. This is due to the limitations of existing authentication systems that cannot proceed without human intervention.

To solve this, global standards organizations like W3C and IETF are collaborating to create dedicated authentication specifications that work safely even in the background, without a human present.

First, the W3C is refining a decentralized identity specification called 'did:atp' through the Agent Trust Protocol Working Group. In simple terms, it is a secure digital ID exclusively for AI agents. By applying post-quantum cryptography that remains resistant even to quantum computers, it prevents others from impersonating or forging the agent's identity.

Simultaneously, the IETF is designing the AAuth specification, an agent authentication protocol. This technology allows AI agents to receive delegated authority asynchronously and complete tasks safely without the user having to wait at their screen to press 'agree.' When an agent sends a natural language-based request for permission, the platform verifies it and issues a token—a proxy authorization—that grants exactly the permissions needed. As a result, the agent can process tasks independently and securely without ever knowing the user’s actual password.

KYA-OS — Cryptographically Verifying That It’s My AI Agent

What is the biggest concern when an agent makes payments or decisions on its own in a network? It is the lack of clarity regarding who is responsible if something goes wrong. There is a need for mechanisms to transparently attribute responsibility when an AI makes an unauthorized payment or accesses sensitive information.

To address this, the 'Know Your Agent - Operating System' (KYA-OS) standard was designed. Led by Keycard, a company founded by Jared Hanson, developer of the authentication tool Passport.js, this technology adds a secure identity layer on top of Anthropic’s Model Context Protocol (MCP).

KYA-OS issues cryptographically signed digital identities to agents and uses Ed25519 key pairs to prove which human user each action originated from.

Below is an example of KYA-OS delegated credentials that authorize and sign specific permissions for an agent.

{
  "issuer": "did:key:z6MkuOwner...",
  "subject": "did:key:z6MktAgent...",
  "delegation": {
    "capabilities": ["mcp:read", "mcp:write"],
    "expires": "2026-12-31T23:59:59Z"
  },
  "proof": {
    "type": "Ed25519Signature2020",
    "signatureValue": "z3h8Ax..."
  }
}

Because these encrypted digital signatures are included, the ultimate responsibility for every action an agent executes can be tracked in an immutable format.

Designing Platforms for a Future Where AI Is the First Customer

For developers and entrepreneurs building web services or platforms, it is time to move beyond designs centered on 'human-manual login.' You must proactively prepare AI-friendly system environments so that AI agents can fulfill their roles while you sleep, utilizing secure cryptographic proxy authorizations and isolated accounts. The trend of AI becoming a bona fide participant in the digital economy is already being proven through specific security standards and platform services. Why not rethink your security and account design with the mindset that the first key customer of your next service might be an AI, not a human?