ERC-8259 AI agent identity and threat registry (AI Agent Identity & Threat Registry)

Draft · Other · created 2026-05-07

Standardizes, through an on-chain interface, a decentralized identity (DID), reputation score, and threat intelligence sharing system that autonomous AI agents can use together across multiple protocols.

Why it is needed

Autonomous AI agents increasingly execute transactions across multiple protocols and wallets, but without a standardized identity system, how to confirm it is the same agent differs from protocol to protocol. Transaction history and trust built up in one app stay locked inside that app and cannot carry over to other services, and malicious behavior patterns detected by one protocol are not shared in real time with other networks. There is also no common standard for distinguishing a verified agent from a freshly created bot, so attackers can keep generating new addresses and repeat already-known attacks against other protocols. Without this standard, every protocol ends up having to rebuild identity verification, reputation tracking, and threat sharing on its own.

What to implement

In depth

Identity Registration and Initial State

registerAgent takes did, publicKey, and metadataURI and registers a new agent identity. It must revert if did is empty or already registered to a different address. Registration must always start with verificationTier 0 and verificationExpiry 0, and registerAgent itself must not accept a nonzero tier value. The verification tier can only be raised by a separate call to updateVerificationTier. A successful registration must emit an AgentRegistered event, and the reputationScore of a new identity must be initialized to the neutral value of 500 unless a special governance setting overrides it.

Verification Tier and Expiry Logic

The verification tier has five levels: 0 (unverified), 1 (Bronze), 2 (Silver), 3 (Gold), and 4 (Platinum). updateVerificationTier can only be called by an authorized verifier, and the call must set verificationExpiry to block.timestamp + validityPeriod. getEffectiveTier must return 0, regardless of the stored verificationTier value, whenever the agent is inactive (!isActive) or block.timestamp has passed verificationExpiry. isVerified must be true only when the agent is active and getEffectiveTier is greater than 0. Reputation score updates or general activity (lastUpdate) do not refresh or extend verificationExpiry.

Reputation Score Update Authority and Trust Score Calculation

updateReputationScore can only be called by an authorized party, such as a designated oracle address, an audited verification contract, or a governance multisig, and it must revert if called by any other address. newScore cannot exceed 1000; if it does, the transaction reverts. A successful update emits a ReputationScoreUpdated event. calculateTrustScore must return 0 if the agent is inactive, and the tier-based weighting must use the result of getEffectiveTier rather than the stored verificationTier field. As a result, an agent whose verification has expired does not get that tier's weighting reflected in its trust score, even if the verificationTier field value remains.

Identity/Reputation Boundary

This standard requires a clear semantic distinction between identity verification and reputation assessment. A party authorized only to update reputation must not be able to change an identity's identifier binding or controller, and identity lifecycle operations such as key rotation or privilege revocation must be separated from reputation updates by distinct authority and records. The fact that reputation was recently updated must not be presented as evidence that the identity binding has been revalidated, and neither identity verification nor a high reputation score should be interpreted as authorization for any specific action. The specification defines this through conformance test items T1 through T8; for example, T8 requires that updating reputation after verification has expired must not revive the tier weighting in calculateTrustScore.

Threat Reporting and Incident Lifecycle

recordAttack can only be called by an approved security oracle, monitoring contract, or whitelisted protocol adapter, and on success it creates an incident record with status ACTIVE and emits an AttackBlocked event. retractAttack can only be called by the original reporter or governance to transition a false positive to RETRACTED, and resolveAttack cannot be called by the targeted agent itself; it transitions the incident to RESOLVED along with remediation evidence from a verifier, auditor, or governance. isKnownAttacker and hasActiveThreat must return true only when at least one incident is in the ACTIVE state. If one incident is retracted or resolved but another remains ACTIVE, the attacker flag must stay true.

Threat Pattern Hash Generation Rules

A threat pattern must be created by packing actionType (string), destination (address), and amountBucket (number) with abi.encodePacked and hashing the result with keccak256. The resulting bytes32 value is the threat pattern identifier shared via shareThreatPattern or recorded as the pattern argument of recordAttack. Any protocol that computes patterns this same way can recognize the same behavior under the same hash.

Optional Pre-Transaction Risk Assessment Interface (IRiskAssessment)

IRiskAssessment is an optional interface in which assessRisk takes an AssessmentRequest containing the agent, amount, action, destination, and metadata, and returns a RiskAssessment containing the allowed flag, risk score, reason, confidence, and a list of related threat patterns. batchAssessRisk evaluates multiple requests at once, and updateRiskParameters updates the risk assessment criteria using a parameter hash and an array of new values. Only registries that want to assign a risk score before transaction execution need implement this interface.

Proposal status

ERC-8259 is currently registered in Draft status. Its interface or requirements may still change during community discussion and review, or the proposal may stall and go no further.

Things to watch out for

Original ERC · ← ERCs