Skip to content
Crypto.world

Research

Toward Web 4.0: Bidirectional Trust between AI Agents and Blockchain

Autonomous AI agents are increasingly being deployed on blockchains, but there is no shared vocabulary for describing what trust guarantees each layer of that interaction actually provides, so each…

Toward Web 4.0: Bidirectional Trust between AI Agents and Blockchain
arXiv
2605.08922
Published
2026-05-09
Authors
Yunfeng Xia, Chao Li, Lei Li, Chenhao Zhang, Li Duan, Runhua Xu, Wei Wang

AI summary

Overview

Research area: The intersection of autonomous AI agents and blockchain systems, framed as the emerging "Web 4.0" paradigm, with emphasis on trust, security, and formal modeling (arXiv category: Cryptography and Security).

Technical level: Intermediate. The paper is a Systematization of Knowledge (SoK) rather than a novel protocol design, but it introduces formal notation (a six-tuple interaction model and eight security properties) and assumes familiarity with Ethereum standards, account abstraction, intents, and ZK proof systems.

One-sentence scope: A structured survey that maps how blockchain supplies trust infrastructure to AI agents and how agents in turn participate in blockchain security, consensus, and governance, evaluated through a formal Agent-Blockchain Interaction Model (ABIM) and a five-dimensional scoring framework across 70 EIPs/ERCs, 20 industry projects, and 118 academic papers.

What This Paper Is About

Autonomous AI agents are increasingly being deployed on blockchains, but there is no shared vocabulary for describing what trust guarantees each layer of that interaction actually provides, so each project invents its own ad hoc security assumptions. The authors argue the relationship is bidirectional and symbiotic: blockchain gives agents identity, permissions, intent execution, and economic incentives (the B→A direction), while agents give blockchain automated auditing, consensus participation, and governance input (the A→B direction), with verifiable computation as the substrate beneath both. The goal is to systematize this design space, formalize the security properties each layer must satisfy, and expose where the ecosystem falls short.

Key Contributions

  1. Formal interaction model (ABIM). A six-tuple ⟨𝒜, ℬ, 𝒱, 𝒫, ℐ, ℰ⟩ capturing agents, blockchain state machine, verification function, permission policy, intent function, and economic incentive function, along with eight layer-specific security properties (verification soundness, identity uniqueness, permission enforcement completeness, solver faithfulness, incentive compatibility, detection completeness, validator verifiability, and bounded influence).
  2. Five-dimensional evaluation framework. Verifiability, Minimality of Trust, Expressiveness, Composability, and Maturity, applied uniformly to EIPs/ERCs, industry projects, and academic proposals.
  3. Cross-cutting synthesis. Joint analysis of 70 Ethereum EIPs/ERCs, ABIM compliance across all layers, and 20 representative industry projects, intended to surface ecosystem-level misalignments that single-layer studies cannot see.
  4. New taxonomies and open problems. A three-dimensional taxonomy (agent autonomy × trust model × operational direction), a threat taxonomy of seven attack categories, and identification of nine concrete open research problems.

Main Findings

  • Agent standards are immature. Of 13 direct AI/agent ERCs, only 2 have reached Final status, which the authors identify as a significant ecosystem gap.
  • The A→B direction is essentially unstudied. Isolated works touch AI in consensus and governance, but no unified security framing treats AI as a first-class protocol-layer actor.
  • Intent architectures lack formal analysis. Only one prior formal analysis of intent architectures exists (reference [12] in the paper).
  • The economics are already large. As of December 2024, the aggregate market valuation of blockchain-based AI agents was estimated at $8.6 billion, and Ante's catalog covers 306 AI-plus-blockchain projects.
  • Security incidents are surfacing. An oracle misconfiguration in the Moonwell protocol caused a $1.78M loss; a pull request for the affected component carried a Co-Authored-By trailer indicating AI tooling involvement, although Moonwell's official post-mortem did not attribute the root cause to AI-generated code.
  • Account abstraction is costly. Creating a SimpleAccount under ERC-4337 costs 381,489 gas, roughly 18 times a standard EOA transfer (21,000 gas).
  • Expressiveness trades against verifiability. Each generation of AI (rule-based scripts, ML-assisted systems, LLM agents) increases expressiveness while decreasing verifiability; a rule-based bot can be formally verified against its specification, whereas a 13B-parameter language model cannot with current proof systems.
  • zkML is feasible but slow at LLM scale. zkCNN proved VGG-16 (15M parameters) in 88.3 seconds with 341 KB proofs; zkLLM was the first to prove inference for a 13B-parameter language model, with full inference at sequence length 2048 taking under 15 minutes, still far too slow for sub-second interactive agents. zkPyTorch reports VGG-16 at 2.2 seconds per image and Llama-3 8B at 150 seconds per token.
  • Auditing detection rates vary widely. The false-negative rate ε for LLM-based security auditing ranges from 0.02 (LLM-SmartAudit on benchmarks) to 0.40 (independent evaluations), a spread that directly affects the B1 detection-completeness property.
  • Compound risks emerge from standard interactions. Security analysis by Qi et al. shows that EIP-7702 interacting with ERC-4337 creates compound attack vectors not present in either standard alone.
  • Seven threat categories, unevenly addressed. Impersonation, privilege escalation, intent deception, inference tampering, solver collusion, agent collusion, and model backdoors; agent collusion and model backdoors are described as "completely unexplored."
  • No DAO enforces an AI voting-power cap. The B3 bounded-influence property is not implemented by any current DAO according to the paper.

Methodology in Plain English

The authors treat the field as an architecture to be mapped rather than a system to be built. They split the agent-blockchain relationship into two directions and then into layers: on the blockchain-to-agent side, an agent must first exist with an identity (A1), then be authorized within limits (A2), then express what it wants via intents (A3), then be economically sustained (A4); on the agent-to-blockchain side, agents observe and audit (B1), then validate (B2), then govern (B3). Running underneath both is a "Trust Foundation" of verifiable computation techniques: zero-knowledge machine learning, optimistic machine learning with fraud proofs, and hardware enclaves (TEEs), ordered from strongest-but-slowest to weakest-but-fastest.

They then write down a formal model of this interaction so that every layer has a stated security property that can be checked against real standards. Each standard, project, and paper is scored on five dimensions, and scores are assigned by the authors from specifications, reference implementations, and third-party audits, with judgment calls resolved in favor of the standard's own stated scope. The synthesis compares coverage against four prior surveys across dimensions such as account abstraction, intent architecture, AI-plus-consensus, and formal modeling.

Note: the supplied text is truncated during Section 3.1, so the detailed treatment of opML, TEE, and the A→B layers (Sections 3.2 through 8) cannot be summarized here beyond what the abstract and introduction state.

Why This Matters

Research impact. The paper gives the field its first shared formal vocabulary for agent-blockchain interactions and the first survey the authors know of to systematize both directions, evaluate all agent-relevant EIPs/ERCs under one framework, and formalize the interaction as a model. It reframes agents as protocol participants rather than mere tools, opening a research agenda around AI in consensus and governance that prior surveys left empty.

Real-world applications:

  • Agent identity and discovery: ERC-8004 registries for identity, reputation, and validation, and ERC-8126 self-registration with ZKP-based privacy and a 0–100 risk score.
  • Verifiable AI inference: ERC-7992 for model registration and ZK inference proof verification, and ERC-7007 for verifiable AI-generated content NFTs.
  • Agent-to-agent commerce: PAN Network combines ERC-8004 with the HTTP-native x402 payment protocol for discovery, price negotiation, and settlement.
  • Bounded delegation for trading and DeFi: session keys and modular accounts under A2, plus intent specification and solver execution under A3, which is where MEV and solver collusion risks concentrate.

Industry relevance. The finding that only 2 of 13 direct AI/agent ERCs are Final tells builders that the standards they depend on are still moving, and that specification maturity does not equal adoption, since many Draft-stage standards are already deployed in production. The five-dimensional scores and the ABIM compliance analysis give protocol designers and auditors a checklist for locating where a given system's trust assumptions actually sit.

Future Directions

  • Mature the agent standards. Push the 11 non-Final AI/agent ERCs toward stability, and reconcile interacting standards whose combination introduces attack vectors absent from either alone (as with EIP-7702 and ERC-4337).
  • Formalize intent architectures. The paper identifies intent systems as lacking formal analysis; the ABIM solver-faithfulness property (A3) supplies a target specification to prove against.
  • Build security framings for AI as a protocol actor. Extend the exploratory B2 and B3 layers, including probabilistic correctness bounds for AI validators (replacing deterministic BFT assumptions) and mechanisms that could enforce the bounded-influence cap τ on AI voting power.
  • Scale verifiable computation to interactive latency. The gap between zkLLM's under-15-minute proving for a 13B-parameter model and the sub-second response interactive agents require defines the central open problem in the Trust Foundation.

The paper itself names nine concrete open problems, whose individual content is beyond the truncated portion of the text provided.

Target Audience

Blockchain protocol designers and standards authors working on account abstraction, agent identity, and intent systems; AI safety and security researchers interested in verifiable computation and LLM-based auditing; smart contract auditors and DeFi risk engineers; and graduate students or technical product leads who need a structured map of the agent-blockchain landscape before diving into individual papers. Readers seeking implementation tutorials or benchmark results for a specific zkML system will find the survey framing useful but will need to follow its 118 cited papers for depth.

Authors’ abstract

Autonomous AI agents are increasingly deployed on blockchain platforms, yet the design space that governs their interaction remains poorly understood. This convergence, where autonomous agents operate on and within decentralized systems, is a defining feature of the emerging Web~4.0 paradigm. This paper presents a Systematization of Knowledge organized around a bidirectional trust framework. In the B $\boldsymbol{\rightarrow}$ A direction, we examine how blockchain provides…

The full abstract on arXiv
Read the original paper