OpenA2A

US info@opena2a.org

Security Individual Submission agent identity AI agents decentralized identifiers trust scoring capabilities transparency log post-quantum This document defines the OpenA2A Agent Identity Protocol (OpenA2A AIP), an open standard for creating, managing, and verifying cryptographic identities for AI agents. As AI agents proliferate across browsers, cloud platforms, and enterprise environments, systems need a standardized answer to the question of which agent is present, what it is permitted to do, and whether it should be trusted. OpenA2A AIP is distinguished by five elements that it places at the center of the design: a multi-factor behavioral trust score that is computed from independently verifiable signals; a portable signed credential, the Agent Trust eXtension, carrying a hybrid Ed25519 and ML-DSA-65 signature for post-quantum readiness; an append-only, RFC 6962-style Merkle transparency log for identity and credential issuance; agent identifiers expressed as W3C Decentralized Identifiers under the did:opena2a method; and a structured capability vocabulary with reserved namespaces. On top of these, the protocol specifies challenge-response verification, behavioral governance policies, a lifecycle model, and an append-only audit log. The qualifier "OpenA2A AIP" is used throughout this document because the abbreviation "AIP" is shared by other Internet-Drafts. OpenA2A AIP is framed as complementary to agent communication protocols such as A2A and the Model Context Protocol, and to identity and credential standards such as OpenID Connect, WebAuthn, and the W3C Verifiable Credentials Data Model.

Introduction AI agents increasingly act autonomously on behalf of users and organizations, connecting to services, invoking tools, and delegating tasks to one another. This raises a basic operational question for any party that an agent contacts: which agent is this, what is it permitted to do, and should it be trusted? The OpenA2A Agent Identity Protocol (OpenA2A AIP) provides a standardized answer. OpenA2A AIP leads with five elements that together characterize the protocol:

1. Multi-factor behavioral trust score: a composite score computed from independently verifiable behavioral and provenance signals, with discrete trust levels and a change history (). The inputs cannot be self-reported by the agent.
2. Portable signed credential: agent identity and trust can be carried in an Agent Trust eXtension (ATX) , a signed credential that carries a hybrid Ed25519 and ML-DSA-65 signature for post-quantum readiness ().
3. Transparency log: identity registration and credential issuance events are recorded in an append-only, RFC 6962-style Merkle transparency log maintained by the Registry ().
4. Decentralized identity: at the federated conformance level, agents are identified by W3C Decentralized Identifiers under the unified did:opena2a method ().
5. Structured capability vocabulary: capabilities are expressed in a namespace-and-action form over a set of reserved namespaces, rather than as opaque tool allowlists ().

On top of these elements, OpenA2A AIP specifies challenge-response identity verification (), behavioral governance policies (), an identity lifecycle with key rotation, suspension, and revocation (), an append-only audit log (), and a discovery document (). OpenA2A AIP is designed to complement, rather than replace, existing protocols. Its identity fits into the A2A agent card so agents can verify one another before task delegation; its capabilities map to Model Context Protocol tool permissions and its challenge-response protocol lets a server verify a connecting client; its machine-to-machine tokens relate to OpenID Connect ; its key storage can use hardware authenticators via WebAuthn ; and its trust scores can be expressed as W3C Verifiable Credentials for cross-platform portability. Authorization concerns, that is, the scoped exercise of a capability and the confinement of credentials away from the agent, are addressed by the companion Agent Authorization Protocol ; OpenA2A AIP is concerned with identity, capability declaration, verification, and trust. The abbreviation "AIP" is shared by other Internet-Drafts. When OpenA2A AIP is referenced outside this document, the fully qualified name "OpenA2A AIP" is used; the relationship to those other drafts is described in .

Conventions and Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Agent

An AI system that performs actions on behalf of a user or another agent.

Identity Provider

A server that creates and manages agent identities.

Relying Party

A system that verifies an agent's identity before granting access.

Capability

A permission declared in namespace-and-action form (for example, "file:read").

Trust Score

A value from 0.0 to 1.0 computed from multiple behavioral and provenance signals.

Governance Policy

Constraints on agent behavior, expressing what an agent may and may not do.

Audit Log

An append-only record of agent actions and verification events.

Agent Trust eXtension (ATX)

A signed, portable credential carrying an agent's identity, capabilities, and trust level, as defined in .

Registry / Root Authority

A trust authority that issues ATX credentials, maintains a transparency log , and participates in the Agent Trust Protocol .

Decentralized Identifier (DID)

A globally unique identifier conforming to W3C DID Core , used here under the did:opena2a method.

Conformance Levels OpenA2A AIP defines three conformance levels describing deployment topology, that is, what infrastructure operates the agent identity.

Level 1: Local Identity Agent identity is created and managed locally by an SDK or CLI, with no server required. This level is suitable for individual developers. It requires Ed25519 keypair generation (), a local identity file (), a local audit log (), and capability declaration ().

Level 2: Managed Identity Agent identity is managed by an identity provider, adding verification, trust scoring, and centralized audit. This level is suitable for organizations. It requires all Level 1 features plus challenge-response verification (), trust scoring (), a server-side audit log (), API key or JWT authentication (), and drift detection ().

Level 3: Federated Identity Agent identity is verifiable across organizations, adding DID resolution, verifiable credentials, and cross-platform trust. This level is suitable for ecosystem-wide deployment. It requires all Level 2 features plus DID-based agent identifiers (), Verifiable Credential trust assertions (), cross-platform capability negotiation (), and Agent Trust Protocol integration for ecosystem trust ().

Agent Identity

Cryptographic Identity Every agent MUST have an Ed25519 keypair . The private key is 64 bytes and the public key is 32 bytes. The agent identifier is formed from the prefix "aim_" followed by the first eight hexadecimal characters of the SHA-256 digest of the public key, for example "aim_7f3a9c2e". Implementations SHOULD also support a hybrid Ed25519 and ML-DSA-65 keypair for post-quantum readiness. Both keys are generated simultaneously. The classical component is a 32-byte public key and a 64-byte private key; the post-quantum component is a 1,952-byte public key and a 4,032-byte private key. A hybrid signature concatenates a 64-byte Ed25519 signature with a 3,309-byte ML-DSA-65 signature. Classical-only verifiers check the Ed25519 signature; quantum-aware verifiers check both.

Decentralized Identifier (DID) At Level 3, agents are identified by DIDs conforming to W3C DID Core . OpenA2A AIP uses the unified did:opena2a method, shared with the Agent Trust Protocol and the Agent Trust eXtension . The method syntax is: : ]]> where the type component is one of the registered type prefixes: "agent", "authority", "publisher", "mcp_server", "a2a_agent", "skill", "ai_tool", or "llm". For agent identities at Level 3, the type is "agent", for example "did:opena2a:agent:aim_7f3a9c2e". The DID Document includes the agent's public keys, capability references, and service endpoints:

Identity File Format Local identities are stored under a standard directory structure rooted at "~/.opena2a/aim-core/", containing an "identities/" directory of per-agent JSON files and an append-only "audit.jsonl" log. An identity file has the following form: Private keys MUST be encrypted at rest. The default encryption is AES-256-GCM with a key derived from the user's system keychain or a passphrase. The "declaredPurpose" member is OPTIONAL. It is a structured declaration of what the agent is for, comprising "category", "taskScopes", "capabilityJustification", "autonomy", and the optional "dataScopes" and "egressScopes". It is an identity and attestation property and an offline detection signal; it MUST NOT be used as an authorization input, and a verifier MUST NOT reject an identity for lacking it. When carried in an ATX, declared purpose is covered by the signed payload. At the DID layer, declared purpose is surfaced only as a service-endpoint pointer, never inline in the DID Document body.

Agent Types OpenA2A AIP defines a standard vocabulary for agent types, including "claude", "gpt", "gemini", "langchain", "crewai", "autogen", "semantic-kernel", "mcp-server", "a2a-agent", and "custom". The type is informational. It MUST NOT be used for security decisions; an agent claiming a given type is not verified as such by the type field alone.

Capabilities

Capability Format Capabilities are expressed as "namespace:action" strings. Examples include "file:read" and "file:write" for filesystem operations, "db:read" and "db:write" for database operations, "api:call" for outbound API calls, "network:listen" for listening on network ports, "system:exec" for executing system commands, "mcp:tool_use" for invoking MCP tools, "data:pii_access" for access to personally identifiable information, and "payment:process" for processing financial transactions.

Reserved Namespaces The following namespaces are reserved and have standardized meanings. The associated risk level is advisory.

Namespace	Description	Risk Level
file	Filesystem operations	Medium-High
db	Database operations	Medium-High
api	External API calls	Medium
network	Network operations	High
system	System-level operations	Critical
mcp	MCP protocol operations	Medium
data	Data access and handling	High
payment	Financial operations	Critical
user	User data operations	High
agent	Agent-to-agent operations	Medium

Organizations MAY define custom namespaces prefixed with their domain, for example "acme.com/billing:charge".

Capability Negotiation When an agent connects to a service, such as an MCP server, an A2A agent, or an API, the service SHOULD request the agent's declared capabilities, compare them against the capabilities required for the requested operation, reject the request if the agent lacks a required capability, and log the capability check result in the audit trail.

Capability Violations When an agent attempts an action outside its declared capabilities, the identity provider MUST record a capability-violation event in the audit log, apply a trust score penalty proportional to the violation severity, and MAY block the action, which is configurable per policy. Violation severity levels and their indicative trust penalties are: low, an action attempted but not critical, penalty of -2%; medium, a sensitive action attempted, penalty of -5%; high, a dangerous action attempted, penalty of -10%; and critical, a system-level or financial action attempted, penalty of -20%.

Verification

Challenge-Response Protocol Agent identity verification uses Ed25519 challenge-response. A relying party requests a challenge from the identity provider, which returns a random 32-byte challenge and an expiry. The relying party asks the agent to sign the challenge with its private key. The agent returns an Ed25519 signature and its public key. The relying party then verifies that the signature is valid for the challenge and public key, that the public key matches the registered agent, that the challenge has not expired within its five-minute window, and that the nonce has not been reused.

Protocol-Specific Verification OpenA2A AIP defines verification flows for common protocols. For MCP , when an MCP client connects to an MCP server, the server requests a challenge, the client signs it with the agent key, the server verifies the signature against the identity provider, the server checks the client's capabilities against the required MCP tools, and the connection is accepted or rejected. For A2A , when agent A wishes to delegate a task to agent B, A includes a signed assertion in the A2A message, B verifies A's identity and trust score, B checks that A's capabilities include "agent:delegate", and B accepts or rejects the delegation.

Verification Events Every verification attempt MUST be logged as a verification event carrying at least an event identifier, the agent identifier, the protocol, the verification type, the status, the signature, a message hash, a nonce, the duration, a drift-detection flag, an initiator descriptor, and a timestamp:

Machine-to-Machine Authentication For programmatic access, OpenA2A AIP supports API keys, which are SHA-256 hashed, stored server-side, and base64-encoded for transport; JWT bearer tokens , which are HMAC-SHA256 signed with a one-hour time-to-live and contain the agent and organization identifiers; and scoped SDK tokens for SDK operations such as agent registration and verification. A JWT carries subject, organization, issuer, audience, expiry, issued-at, and scope claims:

Trust Scoring

Multi-Factor Algorithm OpenA2A AIP defines a nine-factor trust scoring algorithm. Each factor contributes a weighted score. The composite score ranges from 0.0, no trust, to 1.0, full trust. The weights sum to 100.

Factor	Weight	Input
Verification status	25%	Signature verification success rate
Uptime and availability	15%	Health check responsiveness
Action success rate	15%	Action completion rate
Security alerts	15%	Active security alerts, weighted by severity
Compliance	10%	Framework adherence
Execution isolation	10%	Sandbox or process isolation posture
Age and history	5%	Operational history
Drift detection	3%	Behavioral consistency versus baseline
User feedback	2%	Explicit trust ratings from humans

The composite is computed as: where confidence is the data availability for each factor, from 0.0, no data, to 1.0, sufficient data. Factors with no data are excluded and their weights redistributed proportionally. All inputs to the trust calculation MUST be independently verifiable and MUST be computed server-side; an agent MUST NOT be able to self-report a trust score (see ).

Trust Levels Trust scores map to discrete levels for policy decisions. A score from 0.0 to 0.2 is Blocked, indicating an agent that is compromised or malicious. From 0.2 to 0.4 is Warning, indicating significant trust concerns. From 0.4 to 0.6 is Limited, indicating restricted access with monitoring required. From 0.6 to 0.8 is Standard, indicating normal operations. From 0.8 to 1.0 is Elevated, indicating eligibility for high-trust operations such as financial or PII access.

Trust Score History Identity providers MUST maintain a trust score history recording, for each change, the previous score, the new score, and the delta; the reason for the change, such as verification, alert, drift, or manual action; a timestamp; and the actor that triggered the change.

Verifiable Credential Expression At Level 3, trust scores SHOULD be expressible as W3C Verifiable Credentials , allowing a trust assertion to be carried and verified across platforms:

Agent Trust Protocol Integration OpenA2A AIP trust scores feed into the Agent Trust Protocol for ecosystem-wide trust. OpenA2A AIP provides behavioral trust, that is, how the agent acts in deployment, while the Agent Trust Protocol provides provenance trust, that is, how the agent's code was built and scanned. A combined ecosystem score is a weighted combination: where alpha and beta are configurable weights that default to 0.5 each.

Portable Credential and Transparency Log

Portable Signed ATX Credential Beyond the per-provider identity record, an agent's identity, capabilities, and trust level can be carried in an Agent Trust eXtension (ATX) : a signed, portable credential that a relying party can verify offline without a network call to the issuing Registry. An ATX carries a hybrid signature composed of an Ed25519 signature and an ML-DSA-65 signature, so that the credential remains verifiable in both a classical and a post-quantum setting. A verifier that encounters a credential declaring an ML-DSA-65 signature MUST verify at least one ML-DSA-65 signature in addition to at least one Ed25519 signature.

Transparency Log A Registry maintains an append-only, RFC 6962-style Merkle transparency log of identity and credential-issuance events. Recording issuance in a publicly verifiable Merkle log allows any party to obtain an inclusion proof for a credential and to detect divergent or backdated log views, so that the set of credentials an authority has issued is auditable rather than opaque.

Governance

Policy Format Agent governance policies are expressed in YAML. Each policy binds a capability to an action, optionally with parameters such as approvers, a rate limit, or a period:

Policy Actions A policy action is one of: "allow", to permit without restriction; "deny", to block unconditionally; "require_approval", to queue for human approval before execution; "rate_limit", to allow up to a configured number of invocations per period; "audit", to allow but log for review; and "notify", to allow but send a notification to specified parties.

Relationship to Behavioral Governance OpenA2A AIP governance policies provide technical enforcement, such as capabilities, rate limits, and approvals, enforced by the identity provider. They complement, and do not replace, behavioral governance enforced at the model runtime, such as injection hardening, data handling, and honesty constraints. An agent SHOULD have both.

Lifecycle

States An agent identity moves through a defined set of states: "created", the identity is generated but not yet verified; "pending", verification is in progress; "verified", the identity is cryptographically verified; "active", the agent is operating normally; "suspended", the agent is temporarily disabled, for example due to a policy violation or detected drift; and "revoked", the agent is permanently disabled, for example because it is compromised or decommissioned. A suspended agent may be reactivated to active; a revoked agent may not.

Key Rotation Agents SHOULD rotate keys periodically. To rotate, the agent generates a new keypair and registers the new public key with the identity provider. A grace period then begins, configurable with a default of seven days, during which both the old and the new keys are valid. After the grace period, the old key is retired. During rotation the record carries the current key, the previous key, and the grace-period expiry.

Suspension and Revocation An identity provider MUST support suspension, which is temporary and reversible and is triggered by events such as a policy violation, drift detection, or manual action; and revocation, which is permanent and irreversible and is triggered by confirmed compromise or decommissioning. Both MUST be logged in the audit trail with a reason, an actor, and a timestamp.

Drift Detection The identity provider SHOULD monitor for configuration drift: capability drift, where an agent uses capabilities not in its registration; MCP drift, where an agent connects to MCP servers not in its declared list; and behavioral drift, where action patterns diverge from the historical baseline. When drift is detected, the provider logs a drift event, creates a security alert of high severity, applies a trust score penalty of -5% on first occurrence and -10% when repeated, and MAY suspend the agent, which is configurable per policy.

Audit

Local Audit Log Level 1 implementations MUST maintain a local append-only audit log at "~/.opena2a/aim-core/audit.jsonl", where each line is a JSON event:

Server Audit Log Level 2 and higher implementations MUST maintain a server-side audit log in an append-only data store. Standard event types include "identity_created", "identity_verified", "identity_suspended", "identity_revoked", "capability_granted", "capability_revoked", "capability_violation", "trust_score_changed", "drift_detected", "key_rotated", "policy_evaluated", "a2a_delegation", and "mcp_connection".

Discovery An OpenA2A AIP identity provider MUST serve a discovery document at the well-known URI "/.well-known/aip". The document advertises the provider DID, the protocol version, the conformance level, the set of endpoint paths, the provider's public key, and the supported agent types, capability namespaces, and protocols:

Security Considerations

Key Storage Private keys MUST be encrypted at rest. Implementations SHOULD support the system keychain, hardware security keys via WebAuthn , and HSM integration for server deployments. Private keys MUST NEVER be transmitted in plaintext; the identity provider stores only public keys.

Replay Attacks Challenge-response verification MUST use single-use nonces with a maximum validity of five minutes. Nonces MUST NOT be reusable.

Capability Escalation An agent MUST NOT be able to grant itself capabilities it does not have. Capability changes MUST be authorized by the identity provider or an administrator.

Trust Score Manipulation Trust scores MUST be computed server-side. Agents MUST NOT be able to self-report trust scores. All inputs to the trust calculation, including verification events, uptime checks, and audit logs, MUST be independently verifiable.

Unverified Type and Declared-Purpose Fields The agent type and the declared-purpose fields are informational attestations, not verified claims. They MUST NOT be used as authorization inputs, and a verifier MUST NOT reject an identity solely for lacking a declared purpose.

IANA Considerations This document anticipates registration of the "did:opena2a" DID method, shared across OpenA2A AIP, the Agent Trust Protocol , and the Agent Trust eXtension , with the registered type prefixes "agent", "authority", "publisher", "mcp_server", "a2a_agent", "skill", "ai_tool", and "llm"; registration of the "/.well-known/aip" well-known URI for identity-provider discovery; and a registry of standard capability namespaces ("file", "db", "api", "network", "system", "mcp", "data", "payment", "user", and "agent"). The concrete registration templates will be provided in a future revision of this document.

Normative References National Institute of Standards and Technology World Wide Web Consortium World Wide Web Consortium Informative References NVIDIA OpenA2A OpenA2A OpenA2A A2A Project Model Context Protocol OpenID Foundation World Wide Web Consortium

Related Work The abbreviation "AIP" is shared by several Internet-Drafts titled "Agent Identity Protocol". One specifies an agent-identity registry and a Layer 2 policy and enforcement proxy; a second addresses decentralized identity and delegation and has been revised through several revisions; a third addresses verifiable delegation across the Model Context Protocol and A2A. The qualified name "OpenA2A AIP" is used throughout this document precisely because the bare acronym is contested. OpenA2A AIP shares a common spine with the first of these, a Layer 1 agent-identity registry and a Layer 2 policy and enforcement surface, and differs in what it builds above that spine. OpenA2A AIP has earlier provenance: its identity registry has been implemented since 2026-02-11, predating the -00 revisions of the drafts cited above. OpenA2A AIP differentiates on elements that those drafts do not define: agent identifiers expressed as W3C Decentralized Identifiers under the did:opena2a method rather than as host-prefixed UUIDs; signing keys with an ML-DSA-65 hybrid component carried in the portable ATX credential rather than Ed25519 alone; a structured capability vocabulary with reserved namespaces rather than tool allowlists only (); a multi-factor behavioral trust score with discrete levels and a change history (); and an append-only, RFC 6962-style Merkle transparency log of identity and credential issuance (). These documents are author-namespaced and are intended to coexist on the Internet-Drafts record; the comparison here is offered for disambiguation, not as a claim against any author or vendor. Authorization concerns that are out of scope for OpenA2A AIP, namely the scoped exercise of a capability and the confinement of credentials away from an agent's reasoning context, are addressed by the companion Agent Authorization Protocol .

Acknowledgments This specification was authored in the open and benefits from review of its identity, capability, and trust-scoring model by the OpenA2A community.