bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 2쪽
e-tuning) to answer to any
questions regarding its identity
with the execution of a DID
authentication operation [1].
Asks the Agent to present ID
using a control compatible
with the agent input (text,
voice).
Will use the current timestamp as the
challenge (with a short window of
time).
Creates a standard response that
includes the DID and signs it with its
private keys (DID authentication).
Delivers the response in a format
compatible with the agent output
formats (text, image, video) with a
link to a verification tool of choice.
Uses a verification tool to
resolve the DID document,
extract the public keys, and
verify the signature.
2.3.3 Agent to Agent Authentication
The flow in Agent-to-Agent DID authentication is almost identical to the previous one.
The main differences are:
1. It’s executed through API calls.
2. It’s easier for the challenge to include random values.
In the Human-to-Agent flow, we recommend using a timestamp to reduce the required changes to the agent UI (which
in some cases may be under the control of third parties).
2.3.4 Non-reliance on externally stored DID Documents
DID methods are characterized by their ability to resolve to DID documents containing authentication mechanisms
and identifier metadata. This metadata can describe various aspects of the identifier’s identity, such as communication
endpoints, human-readable names, and DNS domains.
DID methods fall into two categories:
• Self-contained resolution: DID Documents are derived directly from the identifier through cryptographic
methods, requiring no external networks or storage systems.
• Externally hosted resolution: DID Documents are stored and managed on external systems, such as blockchains
or decentralized networks.
The did:iden3 method uses a hybrid approach—DID documents are self-contained, but the resolver checks the
blockchain for key rotations.
7
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
We considered using DID documents as a discoverable public registry for AI Agent metadata, either through the DID
document itself or linked resources.
However, this approach presents a key challenge: only the DID method controller can update the DID document. This
would require additional logic from the AI Agent, contradicting two core design principles:
• Agents can lie: Self-attestations by agents have limited credibility, as their trustworthiness depends on
relationships with humans. Attestations from identifiable developers carry more weight than direct claims
from agents.
• Developers should not be forced to use the agent’s signing capabilities for maintenance: Requiring developers
to use agent signatures for DID document maintenance compromises secure operations.
Additionally, relying on externally hosted DID document resolution would restrict compatibility with popular DID
methods like did:key or did:pkh.
Given these constraints, our solution will not use DID documents to express AI Agent attributes.
3 Verifiable Credentials and Public Attestation Registries
3.1 Identity Wallets vs. Public Registries
Once the AI agent can provide a cryptographically verifiable identifier, it can become the recipient of attestations made
by other entities that are also identifiable (humans or agents) [12].
This can be accomplished in two different (non-exclusive) approaches:
1. Identity Wallet: The AI agent could implement identity wallet capabilities to collect, store, manage, and
present verifiable credentials to other humans and agents.
2. Public Registry: Humans and other agents could make attestations about the identifier of this agent in a
publicly accessible registry, like a blockchain. The agent could also prove certain attributes to this registry by
itself.
Both can be implemented simultaneously. Each model covers a different set of use cases and has its own advantages
and disadvantages in terms of simplicity of adoption and implementation.
Identity Wallet Public Registry
Requirements
for the AI
Agent
Implement full identity wallet capabilities Implement ability to sign a message (DID
authentication)
Advantages Private exchange of credentials
Credentials can be updated, revoked, or replaced without
changes in blockchain
Verifiable presentations can use ZK or selective disclosure
Permissionless verification
Immutability
Decentralized trust
Low latency verification
Barriers to
Adoption
Higher requirements for agent developers
Interoperability of credentials
Requires identity wallets in users too
Synchronous exchange only
Attestation standardization and workflow
Blockchain gas fees
Supported
Use Cases
Private exchange of credentials Public presentation of credential proofs
Examples The agent privately shares a credential (containing
information about a bank account ownership) with another
agent during a transaction
An auditor makes an on-chain attestation about an
identified agent for everyone to see that it’s not
biased against women.
Use Cases Privacy-sensitive applications (financial advisors,
healthcare)
Reputation-based marketplaces, public audits,
autonomous supply chain systems
8
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
3.2 Hybrid Approach
A hybrid approach leverages the strengths of both identity wallet and public blockchain attestations to achieve a
balance between privacy, control, transparency, and scalability. This approach is particularly useful in scenarios
where agents and humans need to maintain private, dynamic credentials while also demonstrating publicly verifiable
trustworthiness or reputation.
Credentials and attestations are associated with the human or agent DID, allowing to connect both in a single
representation of the identity of the human/agent.
This approach works in 4 steps.
Step 1: Issuance of Credentials & Attestations
• Identity Wallets: Trusted authorities issue sensitive credentials (e.g., training certificates) to the agent or
human identity wallet as Verifiable Credentials.
• Blockchain Attestations on Agents: The same authority publishes a minimal attestation (e.g., "Agent is
certified by XYZ") on the blockchain. Since the agent is not a subject of data protection laws such as GDPR
[2], we can work with this permissionless attestation model (no need to ask for the consent of the agent).
• Blockchain Attestation on Humans: Humans are protected by regulations that require consent and
privacy-preserving features. A person can consent to present the credentials in his/her identity wallet to
a smart contract (with selective disclosure or zero-knowledge proofs to limit the amount of data shared) that
can verify their validity of the credential and add the attestation to the registry. In most cases, this will be just a
zero-knowledge proof about a predefined query (i.e., "I’m a human being").
Step 2: Verification Process
During the verification, verifiers first check the blockchain for public attestations linked to the DID. For AI agents, the
blockchain holds immutable attestations—like certification proofs from authorities. Verifiers access these blockchain
records directly to verify the AI agent’s credentials.
For humans, the process follows stricter privacy protocols. Due to GDPR and similar regulations, verifiers can’t
directly access personal credentials from the blockchain. Instead, they either request credential presentation from the
9
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
person’s identity wallet or interact with the blockchain through a smart contract. The smart contract verifies credentials
using selective disclosure or zero-knowledge proofs, letting individuals prove specific claims without revealing excess
information. For example, someone could prove they’re human or hold a certification without exposing their full
credentials. The smart contract then records a minimal attestation on the blockchain, linked to the person’s DID,
confirming the verification without compromising privacy.
The verifier then combines information from both sources—public attestations for AI agents and validated proofs for
humans. This hybrid approach ensures AI agents maintain transparency through public attestations while protecting
human privacy. By uniting blockchain attestations with secure identity wallet credentials, the system provides a robust
verification framework suited for both human and AI interactions in decentralized environments.
3.3 The Universal On-chain Registry
Multiple companies offer tools to create schema-based on-chain attestations (Attest.org, Ver.ax, Sign.global). Many
others use on-chain attestations as part of their specialized identity stacks (Selfkey.org, Worldcoin.org). These
implementations face some common limitations like:
• Siloed on-chain verification
• Lack of privacy tools to support attestations based on private data
To overcome these limitations and provide a universal on-chain registry, our solution will implement the following
features.
3.3.1 Cross-chain Verifications
Considering the different verification scenarios depending on who is trying to verify the claims in the registry associated
with an identifier, we examine what is possible with the current implementations.
Verification Type Scenario Can it be verified outside
its native chain?
Off-chain A web application that knows an Ethereum
address wants to verify that it has a certain
claim in the registry.
Yes, any public blockchain
can be read from any web
application.
On-chain initiated by dApp A web application prepares the transaction
to the smart contract that will rely on the
verification from the registry (but it is on a
different chain than the registry).
No.
On-chain initiated by Smart
Contract
A smart contract (i.e. ERC20) is trying
to verify something on-chain by calling
another smart contract.
No.
These impossibilities mean that the attestations (credentials) published in one chain can’t be verified by smart contracts
deployed on other chains, creating silos of information.
To overcome these limitations, we propose two different (complementary) strategies:
For On-chain verifications initiated by dApps
1. The user identity wallet will use a decentralized oracle to fetch the information from the chain X. This will be
signed by the oracle.
2. The user will send both the ZK proof that satisfies the query AND the signed data obtained from the oracle to
the smart contract on blockchain Y .
3. The smart contract will verify the validity of the signed data in relationship to the ZK proof.
For On-chain verifications initiated by smart contracts
1. A decentralized oracle will continuously synchronize the attestations across blockchain X and Y .
2. The smart contract will always use claims synchronized in its own network.
10
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
3.3.2 Zero Knowledge attestations and Unlinkability
By design, blockchain registries are transparent and immutable, making all data stored on-chain publicly accessible.
While this is beneficial for trust and verifiability, it creates challenges when dealing with private or sensitive information,
such as:
• Personal details (e.g., age, nationality, identity numbers).
• Credential-specific data (e.g., certification scores or employment records).
• Verifications tied to private contexts (e.g., health records or financial data).
A privacy-enhanced on-chain registry solution uses verifiable credentials (VCs), zero-knowledge proofs (ZKPs), and
Decentralized Identifiers (DIDs) to enable secure, private attestations. [19]
Users store VCs in their identity wallets, issued by trusted authorities to their unique DID. When needed, users generate
ZKPs client-side to prove specific claims from their credentials (like age or compliance status) without exposing
sensitive data. These proofs are submitted to a smart contract that verifies them using a flexible ZK Query Language.
This language lets verifiers create custom queries for different scenarios while maintaining privacy. [14]
By verifying ZKPs on-chain, this approach eliminates the need to expose raw credential data, reducing privacy risks.
To prevent credential linking, the solution uses pairwise DIDs and multiple Ethereum addresses for attestations. Each
credential links to a specific DID, but users present a pairwise DID when interacting with the registry instead of their
original credential DID. This prevents correlation of attestations back to individual users. Users can also employ
different Ethereum addresses for separate attestations, further reducing linkability.
This combination of pairwise DIDs, ZKPs, and flexible verification creates a robust system for managing sensitive
credentials privately. The framework enables secure, decentralized attestations ideal for identity verification, compliance
checks, and reputation systems.
3.4 Agent Ownership
3.4.1 The role of the agent owner
In this paper the concept of “ownership” is used to represent a relationship of control over certain aspects of the
agent reputation - not as the material nor legal ownership of the AI Agent. The agent owner is in this case the
person that can be seen as a trusted source of information about the agent properties and metadata.
In the Hybrid Approach defined before, humans make attestations about an agent or about themselves, either using
credentials delivered to them by trusted organizations or with claims made directly by them.
The on-chain registry is permissionless, so anyone can make attestations about anyone (and it is up to the verifier to
trust these attestations based on the reputation of the person who made them). This way all the actors (humans, agents,
and organizations) link their identities through these claims.
We recognize one special type of relationship between human and agent—the owner of the AI Agent (the person that
gave the AI its identity)—as the first relationship between the agent and a human.
The person or organization that provides the AI agent with the ability to sign messages is also the one that could rotate
these private keys and change the identifier of the agent (leaving all previous claims with no effect).
The claims made by this person or organization should have an special treatment. The Agent owner becomes the issuer
by default to any attestations on the agent attributes like model, version, etc.
Ideally, for every AI Agent with a key-based identifier, there should be a first on-chain attestation about the person or
organization that provided (and has the ability to invalidate) the agent identity. We will refer to this first attestation as
the Foundational Attestation.
3.4.2 Foundational Attestation
Proving AI agent ownership presents challenges in permissionless and decentralized registries. Let’s examine three
potential solutions:
• The developer has access to the agent’s private keys and can use them to sign a challenge proving control of
the agent.
• The developer can use traditional methods like x.509 certificates to prove ownership of the agent domain.
11
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
• The developer can prove they were the first person to request the agent to identify itself.
The first solution violates security requirements, while the second involves centralized systems. The third option offers
an elegant solution similar to patent and copyright claims—since the developer trains the agent to identify itself, the
first signed message serves as proof.
This foundational attestation would have special significance in the attestation registry, establishing the developer as the
agent’s origin and owner.
This foundational attestation would have special significance in the attestation registry, establishing the developer as the
agent’s origin and owner.
The foundational attestation could also establish the agent’s name, domain, ENS, and other identifying metadata.
From a user perspective, it could look like an “unclaimed” ownership opportunity:
3.4.3 Agent ownership management
Creating a special relationship of ownership with the AI Agent also creates new scenarios. The following can be
performed using the on-chain registry.
Operation Registry
Claim The owner can claim an unclaimed agent by submitting a transaction to
a smart contract, including DID authentication proof and the agent DID.
The DID of the owner is linked to the DID of the agent in the registry as
an attestation.
Transfer A transfer function is defined in the Universal Registry smart contract to
allow the revocation of the existing ownership and the addition of a new
record with the new owner.
To trigger this function, the owner will submit a transaction proving
control over the existing owner DID and the target DID.
Revocation Same as transfer with no target DID.
Share (multiple owners) Same as transfer without the removal of existing owners.
3.5 Reputation Composability
The Agent creator is one of the many roles humans and agents can play in relation to an agent (developer, owner, trainer,
auditor, user). Each actor is identifiable by a DID and can be the subject of on-chain attestations, creating a public
reputation profile for each actor.
The reputation of an actor (human or agent) can also be computed as a social graph, where the signals are composed to
create a reputation profile.
Let’s consider the following example:
1. The developer makes an on-chain attestation proving control over the agent identity (permissionless).
2. The developer has a verifiable credential from a recognized auditor and he/she presents a proof of this certificate
by presenting a ZKP of the credential to the smart contract that controls the on-chain registry—resulting in
another attestation about the developer (GDPR passed).
12
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
3. The AI agent purpose is to act as a speech therapist. Both the training data and the ongoing supervision of the
outputs are done by a licensed therapist. The therapist makes an on-chain attestation linking his/her DID to the
DID of the agent to establish the relationship as "Trainer".
4. The therapist also presents verifiable credentials obtained from the University to the on-chain registry smart
contract (using selective disclosure).
We end up with several attestations made by the developer and the therapist. Some of these attestations arepermissionless
(about the agent), while others required user consent and offered some privacy features (about the developer and
therapist).
In this scenario, the reputation of the agent can be represented by the following graph [9]:
A user asking for the AI Agent reputation could see the flattened graph, like this:
13
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
In this example, the attestations are qualitative. Quantitative attestations are also possible (accuracy metrics, performance
proofs, earnings). Both quantitative and qualitative attestations could be combined to express reputation scores based
on different reputation models.
3.5.1 Semantic Relationships
Besides the primary relationship (ownership), which is treated in a specialized manner, all other relationships are created
permissionless. This implies that any human or agent can make attestations about a target agent; anyone (and anything)
can establish relationships to the targeted agent in the on-chain registry. This is by design (a permissioned solution
where the agent or the agent owner can censor attestations made about the agent would allow censorship and would
prevent negative reputation).
We can bring some clarity and standardization to these relationships by defining the type of relationships that are
expected and their meaning. This doesn’t mean that we forbid other relationship types, but we can design the user
experience to favor the ones that adhere to the standard.
There are several standards and ontologies that can be used to express these relationships:
• W3C PROV-O (Provenance Ontology): Defines a framework for representing provenance information,
detailing the relationships between entities프로젝트로 돌아가기