bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 1쪽
DEEP TRUST: V ERIFIABLE IDENTITIES AND REPUTATION FOR AI
AGENTS
TECHNICAL REPORT
Sebastian Rodriguez, Oleksandr Brezhniev, Dmytro Sukhiy
Privado ID
{srodriguez, oleksandr, dmytro}@privado.id
March 27, 2025
ABSTRACT
The rapid proliferation of artificial intelligence (AI) agents across domains necessitates a robust trust
framework rooted in a multidimensional understanding of AI agent identity. This paper argues that
trust in AI systems must be derived from the relationships these agents establish with other entities
in a reputation system, including humans, organizations, and other AI agents. A trust framework
requires a comprehensive identity model that incorporates architectural, behavioral, legal, and social
dimensions, each contributing to the agent’s reputation within a decentralized ecosystem.
At the foundation of any trust system lies identity, as having verifiable identifiers for agents is
critical for enabling accountability, interoperability, and governance. This paper highlights the
necessity of establishing unique, verifiable identifiers for AI agents, which can anchor reputation
systems and facilitate trustworthy interactions in both human-to-agent (H2A) and agent-to-agent
(A2A) scenarios. Without such identifiers, the reliability of attestations and the integrity of the overall
trust framework cannot be ensured.
We provide a comparative analysis of different identity solutions for AI agents, focusing on their ability
to provide robust identifiers and to express agent attributes. The analysis spans key-based solutions
such as Decentralized Identifiers (DIDs), blockchain addresses, and X.509 certificates, examining
their strengths, limitations, and applicability in decentralized environments. Key challenges such as
protecting private keys from malicious developers, ensuring attribute consistency, and maintaining
identifier validity across changes in agent behavior or architecture are discussed.
This study proposes a hybrid approach that leverages DIDs and public on-chain attestations,
addressing limitations of existing systems. By combining public attestations with privacy-preserving
identity wallets and Verifiable Credentials (VCs), this approach allows agents to maintain trust while
safeguarding sensitive information. Zero-Knowledge Proofs (ZKPs) play a pivotal role in this system,
enabling selective disclosure of credentials and unlinkable attestations. This ensures compliance with
privacy and consent requirements, particularly in scenarios involving human-targeted attestations.
The proposed solution establishes a scalable and privacy-preserving framework for AI agent
identity and trust. By integrating decentralized technologies like DIDs and ZKPs, and supporting
flexible verification methods, this framework provides a future-proof foundation for the responsible
deployment of AI agents in diverse and dynamic ecosystems.
Keywords ai agents, decentralized identifiers, verifiable credentials, self-sovereign identity, zero-knowledge proofs
1 Introduction
As artificial intelligence (AI) agents proliferate across domains from customer support to autonomous driving, their
integration into human-centered and machine-centric ecosystems presents profound opportunities. However, the lack of
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
verifiable identities for these agents hampers trust, restricts accountability, and introduces vulnerabilities in scenarios
where integrity is crucial.
In human-to-agent (H2A) interactions—where users rely on AI for financial advice or medical insights—and
in agent-to-agent (A2A) contexts—where autonomous systems coordinate logistics or execute trades— trust is
indispensable [26].
The concept of trustworthy AI agents is multi-faceted and can be viewed from several perspectives [24], depending on
the stakeholders and the context. Here are some common perspectives [25]:
Perspective Scope How trust is established
Trust the Data Data Collection: Transparency in how data is sourced,
ensuring it is legal, ethical, and unbiased.
Data Quality: Ensuring completeness, accuracy,
consistency, and timeliness of data.
Data Privacy: Respecting user confidentiality and
complying with privacy regulations.
Data Representation: Avoiding bias through
balanced and inclusive datasets.
Third-party audits
Self-attestation
zkML
Trust the training
process
Algorithmic Transparency: Documenting the
methodologies used in training AI systems.
Bias Mitigation: Identifying and eliminating biases
within training datasets and algorithms.
Ethical Oversight: Establishing governance
mechanisms to ensure ethical decisions during the
training phase.
Validation and Testing:Employing rigorous testing
methods to verify accuracy and generalizability.
Explainable AI (XAI): Developing models that can
articulate their learning process and decisions.
Open source models
Third-party audits
Standardized testing
Trusted Execution Environments
zkML
Trust the outcome Accuracy and Precision: Validating that outcomes
are consistent, reliable, and meet predefined goals.
Fairness and Non-Discrimination: Ensuring
decisions are equitable across diverse groups.
Explainability: Providing clear justifications for
decisions to enhance user understanding.
User Feedback: Incorporating feedback to refine
outcomes and maintain relevance.
Auditability: Creating mechanisms for retrospective
analysis of decisions and actions.
Ethicality: Prevent deception and maintain liability
of the AI agent outcomes. [15]
Third-party audits
Standardized testing
Explainable AI
Trust the owner Accountability: Taking responsibility for AI actions,
especially in cases of failure or harm.
Ethical Leadership: Embedding ethical principles in
AI governance and operations.
Compliance and Regulation: Adhering to legal and
industry standards.
User Empowerment: Providing users with
transparency, control, and recourse mechanisms.
Legal ownership
Digital Certificates
IP Protection laws
There is an underlying assumption in all these mechanisms (institutional, software, or hardware-based):
It is possible for an identifiable entity (1) (software, person, organization) to make a verifiable claim (2) about
this particular AI agent (3) [3].
(1) The identification of claim issuers through Self-Sovereign identities is straightforward, with existing market solutions
(such as Privado ID) already implementing this through decentralized identifiers (DID) and verifiable credentials (VC).
2
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
(2) Verifiable data typically relies on cryptographic signatures using asymmetric keys. This includes both Verifiable
Credentials and on-chain attestations signed with private keys corresponding to crypto addresses.
(3) The key challenge in AI agent identity frameworks today lies in precisely identifying “this particular AI Agent.”
This requires meeting three essential conditions:
A. The agent’s identity must have a unique identifier.
B. There must be a mechanism to prove the identifier is solely controlled by the agent, preventing impersonation.
C. The identifier’s validity must reflect the Agent Identity accurately . This raises the “Ship of Theseus”
paradox—when changes in an AI Agent’s data, model, or outcomes alter its identity, should it maintain the
same identifier?
This paper aims to define an identity model for AI Agents that addresses points (1), (2), and (3), establishing the
foundation for all forms of AI reputation discussed above.
2 AI Agent Identity & Identifiers
2.1 AI Identity Framework
Identity comprises the qualities, properties, and characteristics that make an entity recognizable as itself and distinct
from other entities [5]. Since identifiers represent these identities, we must first define "what is the identity of an AI
Agent?" before we can determine what these identifiers represent.
Understanding AI agent identity involves 4 primary approaches:
2.1.1 Architectural Identity
Ensures the identity of the agent is tied to its computational and structural foundation, independent of behavior [25]. Its
key components are:
• Model Architecture: The neural network structure, base LLM, etc.
• Weights/Parameters: Learned parameters that define the model’s function.
• Codebase and Versioning: Source code used to define the agent, including version control.
• Cryptographic Fingerprint: A hash of the model and weights for immutable identification.
• Execution Environment: Runtime attributes like hardware specifications, containers, or Trusted Execution
Environments (TEEs).
The verifiability of these elements is achieved with cryptographic hashing of the models and weights, digital signatures
of the execution logs, and verifiable identifiers like Decentralized Identifiers (DIDs).
These Identifiers (DIDs) are based on the premise that the controller of the DID is in possession of a private key, that
uses to prove its control of the identifier.
The private key is not part of the training data, but is usually managed by external secure systems (TEE, HSM, MPC)
[7] that provide the agent with signature capabilities, which means that they act as possession authentication factors
(something you have) rather than knowledge factors (something that you know).
In this approach, the connection between identifier and AI Agent behavior is not guaranteed: even if the model, data, or
environment changes, the identifier remains valid as long as the AI Agent maintains control over it.
The biggest risk of this approach is that if other entities are making public claims (like certifying the correctness or
security of the agent) by referring to a static identifier, the agent may change its behavior and the claims would still be
valid. We discuss some strategies to reduce this risk, like claim revocation and expiration in section 3.4.
2.1.2 Behavioral Identity
Captures the agent’s dynamic identity, ensuring that its behavior aligns with its architectural identity and intended
design. Its key components are:
• Input-Output Mappings: The relationship between input data and the agent’s responses.
3
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
• Behavioral Signatures: Cryptographic proofs (e.g., zkML) linking behavior to architectural identity. [27]
• Stochastic Elements: Randomness in outputs (e.g., temperature in language models) [8].
• Adaptive Changes: The impact of fine-tuning or reinforcement learning on behavior over time.
The verification methods include behavioral reproducibility tests with controlled inputs and behavioral fingerprints
derived from standardized test cases.
The behavior of an AI agent cannot be fully predicted using just its weights and model, even with constant input. While
weights and the model define the agent’s logic and learned patterns, behavior depends on additional factors such as the
execution environment, hardware precision, and stochastic processes like random initialization or sampling.
Most real-world agents incorporate adaptive mechanisms, randomness, or external decision-making rules, making
behavior emergent rather than static. Thus, the interaction between the model, weights, and runtime context is essential
to understanding and predicting an agent’s behavior accurately.
In this case, the identifier should only be produced and controlled by an agent with specific behavioral characteristics
(something the agent IS). If the agent behavior changes, the agent should lose its ability to prove control over the
identifier.
An example of this would be the use of a zkML based identifier, but this is not feasible today. Zero-knowledge
machine learning (zkML) is advancing but is not yet fully production-ready for large language models (LLMs). The
computational demands of LLMs, combined with the overhead introduced by zero-knowledge proofs, pose significant
hurdles. Current zkML implementations are more suited to smaller models and specific applications [28].
While exact timelines are uncertain, the rapid pace of research and development suggests that zkML could become
feasible for production use with LLMs within the next 3 to 5 years. This projection depends on breakthroughs in
optimizing zero-knowledge proof systems and enhancing the efficiency of zkML frameworks [13, 23].
2.1.3 Legal Identity
The formalized identity of the AI agent as recognized by legal or regulatory frameworks, linking it to its developers,
owners, or operators. Provides a legal basis for trust and accountability, ensuring the agent is tied to human or
organizational entities.
It is composed by:
• Ownership Documentation: Proof of ownership or control (e.g., digital certificates or blockchain attestations).
• Legal Attestations: Certifications or compliance with regulations (e.g., GDPR, ISO standards).
• Accountability Frameworks: Mechanisms that assign responsibility for the agent’s actions to specific entities.
[4]
• Intellectual Property Protections: Legal identifiers for the agent’s code, model, and outputs.
Verification methods include digital certificates issued by trusted authorities, public or private registry entries linking
the agent to its owners or creators and legal contracts or smart contracts tying the agent to its stakeholders.
2.1.4 Social Identity
The perception and reputation of the AI agent, influenced by external entities such as developers, trainers, and its
branding. Establishes trustworthiness and societal acceptance [10] by reflecting external endorsements, trust, and
reputation. Key components are:
• Reputation Signals: Reviews, certifications, or feedback from users and auditors.
• Brand Recognition: Public perception tied to the organization or platform deploying the agent.
• Social Graphs: Attestations or endorsements from other agents or humans (e.g., blockchain attestations,
LinkedIn-style endorsements).
• Transparency: Publicly available information about the agent’s performance, fairness, and ethics.
The verification methods include verifiable credentials stored in an identity wallet, public attestations on a blockchain
or registry, and transparency reports and audits.
4
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
2.1.5 Framework Summary
Category Definition Key Components Verification Methods Identifiers Examples
Architectural
Identity
Inherent, unchanging attributes like
model and weights
Model architecture
Weights
Cryptographic fingerprints
Hashing
DIDs
Execution logs
DID or other PKI-based identifier
Behavioral
Identity
Attributes linked to the agent’s
behavior and interactions
Input-output mappings
Adaptive changes
Stochastic elements
Behavioral tests
Cryptographic logs
Reproducibility proofs
zkML
Legal
Identity
Recognition by legal or regulatory
frameworks
Ownership documentation
Legal attestations
Accountability
Digital certificates
Contracts
Compliance certifications
x.509 certificates
Social
Identity
Reputation and perception based on
external endorsements
Reputation signals
Brand recognition
Transparency
Public attestations
Blockchain credentials
Audits
Relies on other identifiers to receive
reputation signals
2.2 Key-based Identifiers for Architectural Agent Identities
Since the mechanisms to produce identifiers for behavioral agent identities are not ready for production environments,
our solution will define identifiers for architectural agent identities.
The most mature and widely adopted mechanism for providing self-verifiable identifiers is to rely on PKI. Some
examples:
• Self-resolvable DID methods (key-based, decentralized)
• Blockchain addresses (decentralized) [21]
• X.509 certificates (centralized, legally binding)
Using any of these identifiers would require the AI Agent to sign messages with a private key.
While this differs from direct key possession (as the agent could use an MPC network for decentralized signatures or
other mechanisms like HSM modules), all signing mechanisms require an authenticator — essentially a proxy for the
private key that anyone could use to sign.
We must assume developers can act maliciously. This raises the question: "How can we protect agent keys from
developers?"
For simplicity and adoption, we recommend using established security components like HSM, KSM, or TEE to
safeguard the agent’s private keys. This security model choice should remain with individual developers rather than
being standardized. Third parties can audit these decisions as part of the agent’s reputation assessment [16].
However, the solution must prevent developers from accessing or using private keys in any operations related to the
agent’s identity and reputation setup or maintenance. For example:
• Developers should not prove agent ownership by using the agent’s private keys.
• The agent’s keys should not be used for on-chain transactions (such as purchasing ENS tokens), unless the
agent can complete these without exposing its keys.
2.2.1 Digital Signature Algorithms
The generation and verification of key-based DIDs in AI Agents must be standardized and flexible enough to support
multiple implementation strategies. This includes:
1. Support for popular DID methods following W3C standards.
2. Support for multiple formats in the response (text, QR code).
3. The signature verification is flexible - it can happen directly in the client (browser extensions, installed apps)
or through a simple front-end (which could run on decentralized infrastructure like IPFS).
Supporting multiple DID methods opens the door to different signature algorithms and elliptic curves (ECDSA, EdDSA,
RSA).
In order to maximize compatibility with existing systems (Ethereum, Bitcoin), our solution will work with ECDSA on
the secp256k1 elliptic curve. Many projects are already working to provide crypto-wallets to AI Agents, which means
these AI Agents will be able to sign messages and perform DID authentications.
5
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
2.3 Identifier Verifiability: DIDs
Decentralized Identifiers (DIDs) are ideal for AI agents as they offer self-sovereignty, interoperability, and cryptographic
flexibility. Unlike blockchain addresses tied to specific platforms or X.509 certificates reliant on centralized authorities,
DIDs are infrastructure-agnostic, portable, and support key rotation without breaking continuity. They enable
privacy through selective disclosure of credentials, ensuring only necessary information is shared. DIDs operate
on a decentralized trust model, making them resilient to single points of failure, and are cost-effective compared to
blockchain operations or certificate re-issuance. These features make DIDs the most future-proof, adaptable, and secure
identity solution for AI agents in diverse ecosystems [18, 17].
DIDs provide decentralized, self-verifiable identifiers backed by private keys. Using these keys, an AI agent can
generate its own DID and prove control of it to humans or other agents without relying on any third party—all through
a simple signed message.
Some DID methods (like did:iden3) support key rotation, enabling AI agents to update their controlling keys while
maintaining the same DID identifier and preserving all associated claims.
2.3.1 DID Authentication
The DID authentication process is very simple and, for self-resolvable DID methods, it is a peer-to-peer operation that
doesn’t require a connection to any other system.
This can happen in Human-to-Agent or Agent-to-Agent scenarios, allowing for the following differences:
Human to Agent Agent to Agent
Authentication Request User prompt API call
Challenge Generation Timestamp with short time window
or Session ID
Randomly generated
Signature Verification Embedded in the front-end or
linking to external verifier URL
Back-end verification
2.3.2 Human to Agent Authentication
This is an example of how DID authentication can be implemented to authenticate the DID of an agent upon the request
from a user.
6
Verifiable Identities and Reputation for AI Agents TECHNICAL REPORT
User Agent User Experience
The agent has been trained
(fin프로젝트로 돌아가기