AI Agent Identity Certificate (AIC) Extension for X.509 v3

Introduction

Problem Statement Existing X.509 public key certificates, as profiled in , primarily authenticate identity through a cryptographically signed binding between a distinguished name and a public key. Authorization is intentionally left outside the certificate: relying parties perform access control via external IAM policies, OAuth scopes, RBAC databases, or policy engines evaluated after TLS connection establishment. Autonomous AI agents introduce a new requirement that existing certificate profiles do not address. When an agent acts on behalf of a principal, the relying party needs answers to additional questions: Who delegated this authority? What operations are authorized? Under what constraints may the agent operate? How long is the authorization valid? Who remains accountable for the actions performed? These questions cannot be answered by identity alone. They require a standardized mechanism to express delegation relationships, capability constraints, authorization boundaries, and accountability chains within the certificate itself. This document defines an X.509 certificate extension that addresses this gap. The AI Agent Identity Certificate (AIC) extension encodes agent identity, principal binding, capability declarations, delegation modality, authorization boundary constraints, and cryptographic delegation authorization within a single certificate. A companion PrincipalAuthorization extension anchors principal-side grant declarations, authorization constraints, and delegation policies. AIC is designed to complement existing identity infrastructure: SPIFFE/WIMSE identifies workloads via SAN URIs. AIC adds capability containers, principal binding, and session control on top of workload identity. OAuth 2.0 and JWT-based authorization () provides online token exchange. AIC enables offline authorization decisions during the TLS handshake, without external identity provider lookups. Standard X.509 extensions (Certificate Policies, Extended Key Usage) express intended use but not delegation relationships or fine-grained capability constraints. AIC is intended for deployments where: AIC is intended to support deployments that reuse existing PKI infrastructure (enterprise CAs, HSMs, smart cards); Emerging regulatory frameworks require cryptographic traceability of autonomous actions to accountable principals; Offline or air-gapped environments require self-contained certificate authorization without external database lookups; Short-lived workload identities need to be augmented with legal accountability metadata. This document specifies the data model, certificate profile, and validation requirements for AIC. Deployment-specific policy, implementation details, and performance characteristics are outside the scope of this specification.

Scope This document is organized into three categories: Normative (Core): The following sections are normative: Delegation Model, Certificate Profile (Data Model), AIC Extension Definition, Validation Procedure, and PrincipalAuthorization Extension, including their ASN.1 definitions. Implementations MUST conform to these sections to claim AIC compliance. Informative (Profile): The Deployment Models section describes recommended deployment patterns and gateway behavior. These are not required for AIC compliance but represent best practices. Informative (Reference): The Implementation Status section describes a reference implementation. These are provided as implementation guidance and interoperability examples. The core normative content intentionally separates cryptographic delegation from authorization semantics: AIC defines the cryptographic binding between agent and principal, while capability and policy semantics are defined externally.

Design Principles This specification is guided by seven orthogonal concerns, each addressed by a distinct layer: Concern Layer Question Identity AgentIdentity "Who are you?" Authorization PrincipalAuthorization "Who authorizes you?" Capability Capability (Container) "What are you allowed to do?" Delegation DelegationPolicy "How are you allowed to act?" Constraints authorizationConstraints "Under what boundaries?" Enforcement Gateway "Is this request allowed now?" Trust X.509 (PKI) "Is this certificate trustworthy?" Transport TLS "Is this channel secure?" AIC does not redefine trust, transport, or cryptography. It complements X.509 by introducing three orthogonal concepts: Agent Identity, Principal Authorization, and Capability Container. Trust remains in PKI, transport remains in TLS, and enforcement remains in the Gateway. AIC defines the representation and cryptographic binding of authorization-related information (agent identity, principal binding, capability container, constraint container, and delegation evidence). It does not define a universal authorization language: whether an operation is permitted is decided by the capability scheme and the deployment policy, not by the AIC extension itself.

Requirements Language 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. The following terms are used throughout this document:

AIC:: AI Agent Identity Certificate Extension -- the X.509 v3 extension defined in this document.
Agent:: For the purpose of this document, an AI Agent is a software or hardware entity capable of performing actions on behalf of a principal with varying degrees of autonomy.
Principal:: The natural person or organizational entity that authorizes an Agent to act on its behalf. Whether a Principal is legally responsible for actions performed by an Agent is determined by applicable law and deployment policy and is outside the scope of this specification.
Delegation Mode:: The protocol-defined relationship between an Agent and its Principal for attribution and authorization purposes. The interpretation of these modes under applicable law is outside the scope of this specification. In authorized mode, the Agent acts under its own identity with principal consent. In representative mode, the Agent acts on behalf of the Principal within the delegated permission scope.
Capability:: A container for protocol-level operations that an Agent is authorized to perform, defined by a scheme identifier and a capability identifier within that scheme.
Capability Scheme:: A system identified by a scheme identifier (schemeId) that defines the semantics of capabilities. Gateways route capability evaluation to scheme-specific plugins by schemeId.
Credential Bundle:: The certificates presented by an Agent during the TLS handshake, including the agent certificate chain and the principal's certificate.
DelegationAuthorization:: Cryptographic evidence, carried within the AIC extension, that the principal has authorized the Agent. Contains a digital signature over a DelegationAuthTBS structure.
DelegationAuthTBS:: The To-Be-Signed structure whose DER encoding is signed by the principal's private key to produce the DelegationAuthorization.
PrincipalAuthorization:: A companion X.509 extension (OID 1.3.6.1.4.1.66257.1.2) carried in the Principal's certificate, declaring grants, authorization constraints, and delegation policies.
authorizationConstraints:: An OPTIONAL container within AIC and PrincipalAuthorization that defines authorization boundary conditions (IP ranges, concurrency limits, time windows). Constraints are evaluated offline during TLS handshake.
OID:: Object Identifier -- a globally unique sequence of integers used to identify objects in the X.509 standard.
PEN:: Private Enterprise Number -- a globally unique identifier assigned by IANA to organizations for private use in OID space.
SPKI:: Subject Public Key Info -- the ASN.1 structure defined in Section 4.1.2.7 containing the public key algorithm and subject public key.

Related Work Several proposals address agent identity and authorization, either in X.509 certificate extensions or in application-layer protocols. (draft-hood-agtp-agent-cert) defines an X.509 certificate extension that binds an agent identifier and a principal identifier and includes an authority-scope commitment over a set of scope tokens. The scope tokens are carried as a flat list of strings. (draft-sharif-apki-agent-pki) defines five separate X.509 extensions for agent capabilities, delegation, trust scoring, provenance, and behavioral attestation. Each extension is encoded independently. (draft-sharif-x509-agent-identity-profile) defines a single X.509 extension combining a trust level, a list of capability names, a maximum delegation depth, and a kill-switch URI. defines proxy certificates for grid computing. A proxy certificate extends the certificate chain so that the proxy acts on behalf of the issuer; it does not carry agent-specific identity or authorization data. defines attribute certificates as a separate authorization mechanism bound to an identity certificate. Attribute certificates carry attribute/value pairs and are issued by an attribute authority. (arXiv:2603.14332) embeds a hash of a skills manifest in an X.509 extension. Any change to the manifest requires issuing a new certificate. (draft-mishra-oauth-agent-grants) extends OAuth 2.0 with agent grants, using DID-based identifiers and JSON Web Tokens carried at the application layer. It supports multi-level delegation and cascading revocation, and depends on DID resolution. (draft-singla-agent-identity-protocol) defines agent identifiers derived from public keys and a capability manifest carried in DID documents, and depends on DID resolution. WIMSE and SPIFFE define workload identities using SAN URIs (); they identify workloads and do not carry authorization information. This specification follows the X.509 extension approach used by AGTP, APKI, AgentIdentity, and CapabilityBound. It differs in three respects: (1) the authorization data is signed by the principal and the resulting signature is covered by the certificate authority's signature; (2) capabilities use a structured container with a scheme identifier and a capability identifier, so that evaluation can be routed to scheme-specific plugins; and (3) an authorization constraint container carries offline-verifiable boundary conditions. The design goal is to make the authorization decision fully verifiable offline from the certificate and its credential bundle.

Delegation Model

Delegation Modes Two delegation modes are defined: Authorized mode (default): The Agent acts under its own identity. The agent certificate's agentId is recorded as the actor in audit logs, while the principalUid identifies the authorizing principal. The CA evaluates the agent's declared capabilities against the principal's PrincipalAuthorization.grants at issuance time, and the resulting capability set is locked into the certificate. Gateway runtime does not perform a further P_grants superset check for authorized mode. Authorized mode therefore provides snapshot authorization semantics: principal grant changes after issuance do not affect the capability set of an already issued certificate. Representative mode: The Agent acts on behalf of the Principal within the delegated permission scope. The principalUid is recorded as the actor in audit logs, with the agentId recorded as the executor. Representative mode is the exception to the minimal credential bundle model: it requires the principal's certificate and its PrincipalAuthorization extension to be present in the credential bundle presented during the TLS handshake. Deployments that cannot provide the complete bundle MUST use authorized mode. The agent's declared capabilities MUST be a subset of the principal's grants at both issuance time (CA verification) and runtime (gateway verification), as the principal's permissions may change during the agent certificate's lifetime.

Security Envelope Model This specification uses a security-envelope heuristic: reducing either the delegated permission scope or the credential lifetime reduces the potential exposure of a compromised credential. Authorized mode selects narrow scope x longer validity -- the principal selects capabilities at issuance, the capability set is locked into the certificate, and the certificate lifetime is up to 24 hours with automatic renewal. Representative mode selects broad scope x shorter validity -- the agent may exercise the full extent of the principal's permissions (P_grants), but the certificate is short-lived and subject to runtime per-operation P_grants intersection verification.

Permission Intersection The effective permission for any agent operation is governed by the intersection of three sets: The authorization model is the intersection of the principal's grants and the agent's capabilities:

where P_grants is the principal's declared capability grants (in PrincipalAuthorization.grants) and C_agent is the agent's declared capability set (in AIC.capabilities). An operation is authorized only if it belongs to both sets. In authorized mode, C_agent alone serves as the effective capability set (P_grants intersection was already verified by the CA at issuance time and locked into the certificate). In representative mode, the intersection is computed at runtime for each operation. Gateway-local runtime policy (T_policy) -- rate limits, timeouts, routing, and other deployment-side controls -- is an additional enforcement layer applied by the relying party. It is NOT part of the authorization model defined by this specification and MUST NOT be confused with the P (AND) C authorization intersection.

Multi-Level Delegation Single-level delegation (Principal -> Agent, chainDepth = 0) is the default and recommended deployment mode. A delegation chain of depth 1 (Principal -> Agent -> sub-Agent, chainDepth = 1) MAY be supported when a deployment requires it, in which case the delegating agent signs a DelegationAuthorization for the sub-agent: chainDepth = 0 (default): direct delegation from the principal to the agent (Principal -> Agent). chainDepth = 1 (optional): one additional level (Principal -> Agent -> sub-Agent), in which the delegating agent signs a DelegationAuthorization for the sub-agent. Each level of the chain produces an independent DelegationAuthorization signed by the delegator's private key, and capabilities are recursively intersected along the chain. Chain depth is carried in a DelegationDepthControl extension with chainDepth and maxDepth fields. In any chain, chainDepth MUST NOT exceed maxDepth. Chains deeper than 1 are not recommended. As a best practice, deployments SHOULD NOT permit chainDepth > 1: each additional level increases attribution ambiguity (the legal status of intermediate agents and the audit actor semantics), expands the attack surface, and grows the credential bundle. The accountability model of this specification is anchored to the natural person at the top of the chain and is best preserved by limiting delegation to the depths described above.

Certificate Profile (Data Model) A certificate carrying the AIC extension encodes the following abstract data model: Component Description Cardinality Agent Identity Identifier of the agent Mandatory Principal Binding Link to the accountable principal Mandatory Delegation Mode Authorized or representative Mandatory Capability Set Declared operational capabilities Optional Authorization Constraints Offline-verifiable boundary conditions Optional Delegation Authorization Cryptographic evidence of principal consent Mandatory Extensions Vendor-specific or future extensions Optional This data model is encoded as a set of X.509v3 certificate extensions, defined in the following section.

AIC Extension Definition

OID Tree

Capability Glob Matching Syntax Capability matching uses the following matching rules: Pattern Meaning Example scheme:method:path Exact match http:GET:/api/v1/users scheme:method:path/* Single-segment wildcard (no "/") http:GET:/api/v1/* scheme:method:path/** Multi-segment wildcard (crosses "/") http:GET:/api/v1/** scheme:*:path Method wildcard http:*:/api/v1/* scheme:* Scheme-wide wildcard http:* Matching priority (highest to lowest): 1. Exact match 2. Single-segment wildcard 3. Multi-segment wildcard 4. Method wildcard 5. Scheme-wide wildcard When multiple rules match, the highest priority rule applies. If no rule matches, the capability MUST be treated as Deny. Unqualified wildcards (bare * without a preceding namespace and colon separator) are NOT permitted. All wildcards MUST be scoped within a schemeId namespace.

ASN.1 Module This section defines the ASN.1 module for the AIC extension, using the conventions specified in . Encoding tag conventions: Mandatory fields use universal tags (INTEGER, UTF8String, OCTET STRING, SEQUENCE); no context-specific tags are applied. All OPTIONAL fields use context-specific explicit tags ([n] EXPLICIT), numbered from 0 in field order within each structure. Tag numbers are unique within a structure and MAY be reused across structures. Standard external types (AlgorithmIdentifier) follow RFC 5280 encoding conventions. DEFAULT fields MAY be omitted from DER encodings when equal to the default value; implementations MUST accept both forms. The AIC and PrincipalAuthorization extensions MUST be non-critical so that systems unaware of these extensions can safely ignore them.

Example AIC Extension Encoding The following example illustrates the AIC extension for an agent with agentId "agent-1", a principal identified as "zhangsan" in the "corp.com" realm, and a single HTTP capability:

}, capabilities { { schemeId "http", capabilityId "GET:/api/v1/users" } }, delegationAuthorization { reason { reasonCode "SCHEDULED_MAINTENANCE", description "temporary maintenance window" }, requestedLifetime 3600, timestamp "2026-08-18T00:00:00Z", nonce <32-byte CSPRNG value>, signatureAlgorithm ecdsa-with-SHA256, signatureValue } } ]]> In the DER encoding of this example, the DEFAULT fields (version, delegationMode) are omitted as permitted by the ASN.1 module, and the OPTIONAL fields (authorizationConstraints, extensions) are absent. Each field is encoded in the order given in the ASN.1 module using DER rules RFC5280. The signatureValue is produced by the principal's private key over the DER encoding of the corresponding DelegationAuthTBS structure and varies per authorization.

Field Definitions

agentId The agentId field contains a globally unique identifier for the AI Agent instance. The identifier SHOULD be stable across certificate renewals for the same agent instance. Implementations MAY use UUIDs, URN-based identifiers, or DNS-anchored names.

delegationMode The delegationMode field defines the protocol-defined relationship between the Agent and its Principal for attribution and authorization purposes:

authorized (0):: The Agent acts under its own cryptographic identity with explicit principal authorization. The audit trail records the agentId as the actor. This is the default mode.
representative (1):: The Agent fully represents the Principal. Actions are attributed to the Principal (principalUid) in the audit trail. This mode MUST only be permitted when the Principal's certificate explicitly allows representative delegation (allowedMode=representativeAllowed in PrincipalAuthorization.delegationPolicy).

principalUid The principalUid field identifies the natural person or organizational entity that authorizes the Agent's actions. It is encoded as an ASN.1 SEQUENCE:

Where: realm is a globally unique namespace (e.g., an organization domain), limited to 128 characters; identifier is a unique identifier within the realm, limited to 256 characters; keyHash is the hash of the SubjectPublicKeyInfo (SPKI) of the principal's certificate computed with hashAlgo: keyHash = hashAlgo(SPKI). The length is determined by the hash algorithm and MUST NOT exceed 64 bytes (SHA-256 = 32 bytes, SHA-384 = 48 bytes, SHA-512/SHA3-512 = 64 bytes, SM3 = 32 bytes). Using the SPKI hash rather than a certificate fingerprint allows principal certificate renewal without invalidating existing agent authorizations, provided the same key pair is used. hashAlgo identifies the hash algorithm used for keyHash; when omitted, SHA-256 (OID 2.16.840.1.101.3.4.2.1) is assumed. Only hash algorithms with output length not exceeding 64 bytes are supported. The human-readable form {realm}:{identifier}:{keyFingerprint} (where keyFingerprint is the base64url encoding of keyHash per Section 5, without padding) is RECOMMENDED for display and logging purposes only. Machine-level comparison MUST use ASN.1 structural comparison, not string parsing. keyHash MAY be used as a management index (e.g., for cascading revocation, audit, or certificate lookup by principal). Such lookups are for management association only; authorization binding and revocation decisions remain based on keyHash and certificate chain validation.

capabilities The capabilities field is a pure container -- this specification defines only the encoding of capabilities, not their semantics. Capability semantics are entirely defined by the capability scheme identified by schemeId. AIC implementations MUST NOT assign semantics to unknown schemes. Each capability entry is identified by a schemeId that indicates the governing scheme, a capabilityId within that scheme, and OPTIONAL parameters. The capabilityId field supports wildcard characters for glob-based matching: * matches a single path segment (no colon), ** matches across path segments (any depth). Gateways route capability evaluation by looking up registered plugins by schemeId. When the capability required by the current request references an unknown scheme or an unknown capability, the request MUST be treated as Deny (fail-closed). Capabilities carried in the certificate that are not relevant to the current request do not affect the decision.

authorizationConstraints The authorizationConstraints field (OPTIONAL) defines authorization boundary conditions for the Agent. Each constraint reuses the Capability container. The schemeId MUST be one of "varwof/constraint-v1"; any other schemeId MUST be rejected. capabilityId distinguishes the constraint type. The following constraint types are defined as examples; additional constraint types are registered through the capability scheme registry (Section IANA Considerations) and do not require a change to the ASN.1 module: capabilityId parameters Format Description allowed-cidr ["10.0.0.0/8", "192.168.0.0/16"] Allowed IP ranges max-concurrent {"max": 5} Maximum concurrent Agent instances time-window {"start": "22:00", "end": "06:00"} Allowed execution time window (UTC) Constraint semantics MUST define the reference clock, time scale, timezone interpretation, boundary conditions, and behavior under clock uncertainty (e.g., for time-window). Constraints are evaluated with AND logic: all constraints MUST be satisfied for the connection to be accepted. The maximum number of constraints is 32, with a single constraint's parameters field limited to 512 bytes. Unknown constraint types are logged as audit warnings and ignored by default (forward-compatible); strict rejection behavior may be configured by the deployment. Design principle: authorizationConstraints carry boundary conditions determined by the authorizing party, varying per-principal, and changing infrequently. Runtime policies (timeouts, retries, rate limits, routing) are NOT carried in authorizationConstraints and remain in gateway-local policy configuration. The boundary between authorizationConstraints and gateway runtime policy is fixed: authorizationConstraints carry authorization boundary conditions set by the authorizing party (IP ranges, concurrency limits, time windows) that are verifiable offline. Gateway runtime policy (timeouts, retries, rate limits, routing, logging levels) is deployment-specific and MUST NOT be carried in authorizationConstraints. Constraint types registered with the same capabilityId and parameters MUST be interpreted consistently by gateways that implement them. Gateways that do not implement a constraint type treat it as unknown and follow the unknown-constraint handling described above. The constraints within AIC.authorizationConstraints are independent of constraints within PrincipalAuthorization.authorizationConstraints -- PA constraints limit the authorization behavior, AIC constraints limit the execution behavior. Both are checked independently during validation.

delegationAuthorization The delegationAuthorization field provides cryptographic evidence that the principal has authorized this Agent. It contains:

The signature algorithm and signature value are kept as two flat fields -- signatureAlgorithm preceding signatureValue -- following the X.509 certificate convention (RFC 5280 Section 4.1), where signatureAlgorithm precedes signatureValue in the Certificate structure, enabling parsers to determine the algorithm before reading the signature bytes. They MUST NOT be merged into a nested SEQUENCE. reason (mandatory): The reason for this delegation, carried both here and in DelegationAuthTBS (where it is covered by the principal's signature). reasonCode is a controlled vocabulary value in SCREAMING_SNAKE style (e.g., SCHEDULED_MAINTENANCE, AUTO_RENEWAL); description is a human-readable explanation. Both fields MUST be present and non-empty. reason does not appear in PrincipalAuthorization. requestedLifetime: The principal-requested certificate lifetime in seconds. The wire value MUST be in the range 1-86400 (SHOULD 3600-86400). The CA determines the actual NotAfter as min(requestedLifetime, local policy maximum). timestamp: The time at which the authorization was granted. MUST be UTC encoded as GeneralizedTime in Z form. nonce: 32-byte CSPRNG-generated random value for replay protection. The CA verifies nonce uniqueness at issuance time and persists it to prevent reuse. signatureAlgorithm: The algorithm identifier for the principal's signature (e.g., ecdsa-with-SHA256). signatureValue: The DER-encoded signature value produced by the principal's private key over the DER encoding of DelegationAuthTBS. The principal's private key signs the DER encoding of the following DelegationAuthTBS structure:

All ten fields MUST be present in the DER encoding of DelegationAuthTBS. The authorizationConstraints field uses context-specific tag [0] EXPLICIT encoding; its OPTIONAL nature does not break backward compatibility with certificates that do not carry constraints.

Extensions and Extensibility The extensions field in AIC provides an extensibility mechanism for vendor-specific and user-specific metadata, analogous to X.509 v3 extensions but scoped to the AIC context. Each extension entry is identified by a globally unique OID. Vendor-specific extensions SHOULD use OIDs under the vendor's own IANA Private Enterprise Number. Unknown extensions with critical = TRUE MUST cause certificate rejection. Unknown extensions with critical = FALSE (default) MAY be silently ignored.

Certificate Size Constraints Certificates MUST observe the following size limits: Full-protocol safety limit: 12 KB recommended, 16 KB hard limit. This is the DER-encoded certificate size compatible with TCP/HTTP/ DTLS/QUIC gateways. 16 KB corresponds to the QUIC CRYPTO_BUFFER_EXCEEDED limit of the QUIC stacks used by the reference implementation; exceeding it fails the handshake. Handshake certificate (AIC lightweight): 8 KB recommended, 16 KB hard limit. Used for mTLS/DTLS/QUIC handshake. In dual-certificate deployments it contains agentId + principalUid + delegationMode DelegationAuthorization, plus at least one capability entry (the AIC capability set MUST NOT be empty; see Section 4.5.4). The dual-certificate deployment model is an optional deployment profile and does not define a new TLS certificate type. Full authorization certificate: 64 KB recommended, 128 KB hard limit. Carries all capabilities, authorizationConstraints, and extensions; transmitted at the application layer after the handshake. Certificates exceeding 128 KB MUST be rejected. Unlike QUIC, TCP/TLS transports do not impose the 16 KB handshake limit: certificate messages are segmented by the TLS record layer, so larger certificates can be carried in practice up to the hard limit. The number of entries in the extensions field SHOULD NOT exceed 32. The ASN.1 module bounds the capabilities sequence with SIZE(1..MAX) and does not fix a numeric limit on the wire format. Certificates whose capability entries exceed 256 MUST be rejected to prevent DoS attacks; typical deployments carry far fewer entries, and deployments SHOULD keep the capability set as small as practical.

Capability Parameters Intersection When matching capabilities across P_grants and C_agent, parameter intersection follows these rules: If C_agent.parameters exceed the boundary defined in P_grants.parameters, the capability entry is invalid (filtered or rejected). Otherwise, C_agent.parameters are used in full (agent-specified values within principal-defined boundaries). Examples: P_grants C_agent Result max_rows=1000 max_rows=100 Accept, use max_rows=100 max_rows=1000 max_rows=5000 Reject (exceeds boundary)

Validation Procedure This section defines the validation procedure that a gateway or relying party MUST perform when processing an AIC-bearing certificate.

Validation Pipeline The gateway performs the following validation steps in order after TLS handshake completion: Certificate Chain (): Verify the certificate chain up to a trusted root, including signature validation, validity periods, and basic constraints. Revocation Status: Verify that neither the end-entity certificate nor any intermediate certificate in the chain has been revoked. Implementations MAY use CRLs (), OCSP (), or OCSP Must-Staple (). In offline deployments, a locally cached CRL or a short validity window MAY serve as an alternative revocation mechanism. AIC Extension Parsing: Parse the AIC SEQUENCE. If the extension is marked critical and cannot be parsed, the certificate MUST be rejected. DelegationAuthorization Verification: Verify the principal's signature in delegationAuthorization.signatureValue: Reconstruct the DER encoding of DelegationAuthTBS (including reason and authorizationConstraints if present). The TBS is DER-encoded in the field order defined in the ASN.1 module; DEFAULT fields MAY be omitted and are interpreted with their default values. Use the principal's public key (identified by principalUid.keyHash) to verify signatureValue. Cross-verify that hashAlgo(principal_cert.SPKI) equals principalUid.keyHash (hashAlgo omitted means SHA-256). Verify that the certificate validity period (notAfter - notBefore) does not exceed delegationAuthorization.requestedLifetime, and that requestedLifetime is at most 86400 seconds (1 day). Verify that signatureAlgorithm is one of the permitted algorithms: ECDSA with SHA-256 (MUST), ECDSA with SHA-384 or SHA-512 (MAY), RSA with SHA-256 (MUST), RSA with SHA-384 or SHA-512 (MAY), RSA-PSS with SHA-256, SHA-384, or SHA-512 (MAY), or Ed25519 (MAY). Any other algorithm MUST be rejected. PA.authorizationConstraints Check: If the PrincipalAuthorization extension in the principal's certificate contains authorizationConstraints, evaluate them. PA constraints limit the authorization behavior and are independent of AIC.authorizationConstraints. AIC.authorizationConstraints Check: If AIC.authorizationConstraints is present, evaluate each constraint it recognizes with AND logic. All constraints MUST be satisfied. Unknown constraint types are logged as audit warnings and ignored (forward-compatible). This step is executed before capability evaluation for fast rejection of unauthorized connections. Delegation Mode Check: If delegationMode is representative: Load the principal's certificate and its PrincipalAuthorization extension. Verify allowedMode permits representative delegation. Verify that C_agent is a subset of P_grants (P (AND) C (AND) T). Delegation Depth Check: If a DelegationDepthControl extension is present, verify that chainDepth does not exceed maxDepth. As a best practice, deployments SHOULD NOT accept chains with chainDepth > 1 (or maxDepth > 1); such chains are outside the recommended deployment envelope of this specification. Capability Evaluation: For the capability required by the current request, route evaluation to the plugin registered for its schemeId. If no plugin is registered or the capability is unknown, the request MUST be denied. Other capabilities carried in the certificate but not relevant to the current request do not affect the decision. Decision: If all validation steps pass, the relying party SHOULD allow the connection. If any step fails, the relying party MUST reject the request. Deployments SHOULD record the rejection with sufficient diagnostic information for audit purposes.

Offline Validation In offline or air-gapped deployments without access to CRL or OCSP responders, the relying party MUST accept the risk that a revoked certificate may be accepted until the next cache refresh. Mitigations include: Short certificate validity windows (RECOMMENDED <= 1 hour); Locally cached CRL shards with freshness checks; authorizationConstraints evaluated entirely offline during TLS handshake. The principal's certificate is obtained from the Credential Bundle presented by the Agent during TLS handshake (TLS 1.3 allows multiple CertificateEntry messages). If the principal's certificate is not present in the chain and no local cache is available, the gateway MUST reject the connection (Fail-Close). The transport mechanism for the credential bundle is a deployment detail and is not defined as a new TLS message type by this specification.

PrincipalAuthorization Extension The PrincipalAuthorization extension (OID 1.3.6.1.4.1.66257.1.2) is a companion X.509 extension carried in the Principal's certificate. It declares the Principal's authority boundaries, including grant declarations, authorization constraints, and delegation policies.

ASN.1 Definition

The grants field defines the set of capabilities the principal may grant to agents -- serving as the upper bound (P_grants) in the permission intersection model. The authorizationConstraints field carries principal-level authorization boundary constraints, reusing the Capability container with schemeId="varwof/constraint-v1". PA constraints and AIC constraints are evaluated independently -- they operate at different semantic layers and are not in a subset relationship. The delegationPolicy field defines the Principal's delegation boundary: how many concurrent agents are permitted (maxAgents), which delegation modes are allowed (allowedMode), and optionally the maximum session hours (maxSessionHours). maxAgents (in DelegationPolicy) and the max-concurrent constraint (in authorizationConstraints) have distinct semantics: maxAgents limits how many Agent instances the Principal may delegate to concurrently, whereas max-concurrent limits how many concurrent connections a single Agent may establish. They are enforced at different layers (policy boundary vs execution boundary) and are not interchangeable.

Principal Key Rotation When the principal's private key is compromised and a new key pair is generated, the SPKI hash in principalUid.keyHash changes, causing all existing representative-mode agent certificates to fail authorization checks. This is an intentional design: when the key pair changes, all associated agent certificates are automatically invalidated through a key management operation rather than a revocation broadcast. When the principal renews their certificate using the same key pair, the SPKI remains unchanged and existing agent authorizations continue without requiring re-issuance.

Deployment Models

Enterprise/Regulated Model The principalUid field MUST be populated with a verifiable natural person identifier. The delegationAuthorization field MUST be present and cryptographically validated. Representative mode MUST only be used with explicit PrincipalAuthorization delegation grants. Certificate validity SHOULD be short-lived (hours to days). authorizationConstraints SHOULD carry IP range, concurrency, and time window boundaries. All AIC-bearing connections SHOULD be logged.

Consumer/Individual Model The principalUid MAY be omitted or contain a pseudonymous identifier. The delegationAuthorization field MAY be omitted for low-risk operations. Representative mode SHOULD NOT be used. Deployments in this model prioritize consumer privacy protections over granular identity binding.

Extensibility Framework The extensions field in AIC provides an extensibility mechanism for vendor-specific and user-specific metadata, analogous to X.509 v3 extensions but scoped to the AIC context. Each extension in the extensions sequence is identified by a globally unique OID. The following guidelines apply: Vendor-specific extensions SHOULD use OIDs under the vendor's own IANA Private Enterprise Number. Unknown extensions with critical = TRUE MUST cause certificate rejection. Unknown extensions with critical = FALSE (default) MAY be silently ignored. This document specifies the container structure only. Semantics are defined by the OID assignee.

Privacy Considerations (GDPR / Right to Erasure) AIC certificates are static X.509 data objects. Once issued, the principalUid field cannot be modified or deleted from distributed certificates. Deployments MUST ensure that: The identifier field within principalUid does not contain raw personally identifiable information (PII) such as full names, national ID numbers, or email addresses. Internal UUIDs or pseudonymous identifiers MUST be used instead. The mapping table between pseudonymous identifiers and natural persons is stored in a compliant database that supports deletion under GDPR Article 17 or equivalent regulations. Certificate revocation (CRL/OCSP) is the only mechanism available to signal that a principal no longer authorizes a given agent; revocation does not erase the certificate itself. AIC certificates SHOULD be treated as stateless credentials by relying parties: they are parsed and verified during the TLS handshake, after which the raw certificate MAY be discarded and need not be persistently stored. Audit logs SHOULD be minimized: they need not record the full principalUid structure or extension content. A session-bound fingerprint, the operation summary, the decision, and the pseudonymous identifier are sufficient for accountability. Log retention MUST be configurable by the deployment according to applicable jurisdictional requirements. A layered model is recommended: long-term archival for high-accountability regimes, limited cleanup periods for standard deployments, and an ephemeral no-mapping mode for consumer scenarios. Cryptographic evidence and log mappings are independent: the DelegationAuthorization signature can be verified on its own without relying on audit log mappings, so deleting a pseudonym mapping table does not invalidate the cryptographic evidence of authorization. Since certificates are immutable once issued, data-subject rights (access, rectification, erasure) are realized through certificate revocation and re-issuance: revoke the certificate carrying the old identifier, issue a new certificate with a fresh pseudonymous identifier, and delete the associated mapping entries. This specification provides privacy-preserving mechanisms only. Legal compliance (lawful basis, cross-border transfer, breach notification, and other obligations) is the responsibility of the deployment acting as data controller or processor. All data carried in the AIC extension is visible during the TLS handshake. Implementations MUST NOT place sensitive data in any certificate extension.

Security Considerations The AIC extension is not an authentication requirement by itself: the X.509 certificate chain already authenticates the agent, and the AIC extension carries authorization-related information that the relying party may evaluate. Deployments that require AIC-based authorization MUST configure the relying party to require and process the AIC extension. AIC does not constrain the authority of a CA to issue PrincipalAuthorization extensions. The trustworthiness of a PrincipalAuthorization depends on the CA issuance policy and the trust anchor configuration.

Threat Model This specification considers the following adversary models: Network Attacker (E): A network-level adversary capable of intercepting, modifying, and replaying TLS connections. This adversary does not have access to any private keys. Malicious Agent (M): An authenticated agent in possession of a valid AIC certificate that attempts to escalate privileges, impersonate other agents, or exceed its authorized capabilities. Compromised Principal (P): A principal whose private key has been leaked. The attacker can sign arbitrary DelegationAuthorization payloads using the compromised key. Compromised CA (C): An adversary with access to a CA signing key. This adversary can issue arbitrary certificates and is the most powerful in the model.

Threat Mitigations Threat Mitigation Mechanism Agent impersonation Cryptographic identity binding X.509 CA-issued certificate, BasicConstraints CA:FALSE Principal denial of authorization Digital signature on authorization DelegationAuthorization.signatureValue over DelegationAuthTBS Capability escalation Permission subset check P_grants (AND) C_agent (AND) T_policy intersection Cross-CA role spoofing Trust anchor hash verification Trust anchor fingerprint comparison in offline plugin parameters Unknown extension bypass Critical flag enforcement Reject on unknown critical extension Signature replay Nonce binding in TBS 32-byte nonce in DelegationAuthTBS, CA uniqueness check Authorization forgery Dual-signature nesting CA signature over TBSCertificate + principal signature in AIC extension

Offline Validation Risks In offline deployments, the risk of accepting a revoked certificate exists until the next cache refresh. Mitigations include short certificate validity windows and strict authorizationConstraints that limit the blast radius (IP range, concurrency caps).

Authorization Constraint Integrity The authorizationConstraints field is carried inside the AIC extension, which is covered by the CA signature over the TBSCertificate; a modification of any constraint therefore fails certificate signature validation. The effectiveness of constraint enforcement also depends on the gateway evaluating the constraints it recognizes and on constraint plugins implementing the declared semantics. Constraints are boundary conditions that limit the blast radius of a compromised or malicious agent; they are not a substitute for gateway-local policy controls such as rate limiting or resource quotas.

Principal Key Hash Binding The principalUid.keyHash field binds the delegation authorization to the principal's public key through the hash algorithm identified by hashAlgo. The hash output is limited to 64 bytes by the ASN.1 SIZE constraint, and a collision in a cryptographically secure hash function is assumed to be computationally infeasible under the security assumptions of the selected hash algorithm. In addition, keyHash is used to locate and bind the principal's certificate within the credential bundle; signature verification uses the principal's public key after certificate chain validation, so a hash collision alone would not enable an attacker to produce a valid principal signature. keyHash itself does not establish trust: trust derives from certificate chain validation and the principal's signature. keyHash is a locator and binding mechanism only.

Nonce Replay Protection The 32-byte nonce in DelegationAuthTBS is signed as part of the DER encoding, making it a cryptographic input to the principal's signature -- not a plaintext parameter. The CA verifies nonce uniqueness at issuance time and persists used nonces. The gateway MAY optionally maintain a local NonceCache for additional replay detection, but this is not required: the primary defense is the CA-level uniqueness check. Replay protection is therefore enforced at the CA issuance layer (nonce uniqueness at issuance time), not at the gateway runtime; the gateway's optional cache is a secondary defense only.

IANA Considerations

Private Enterprise Number Assignment This document uses the IANA Private Enterprise Number 66257, assigned to Varwof PKI (contact: Jijie Wei, pki@varwof.com). The OID prefix is:

OID Registration The following OIDs are defined in this document:

1.3.6.1.4.1.66257.1.1:: AIC Extension (Section AIC Extension Definition)
1.3.6.1.4.1.66257.1.2:: PrincipalAuthorization Extension (Section PrincipalAuthorization)
1.3.6.1.4.1.66257.1.1.4:: DelegationDepthControl (Section Multi-Level Delegation)

Additional OIDs under branch 3 (National/Industry Certification) are reserved for future allocation. The full OID tree is defined in Section OID Tree.

Capability Scheme Registry This version of the specification does not request an IANA registry. Capability scheme identifiers and constraint types are registered through an external or community registry of scheme identifiers, which follows the vendor/product naming convention. An IANA registry may be requested in a future revision if the scheme namespace is transitioned to IANA administration. The initial entry in the external registry: Scheme Identifier Description Reference varwof/constraint-v1 Authorization boundary constraint (allowed-cidr, max-concurrent, time-window) This document The http and database scheme identifiers are reserved as examples of externally defined capability schemes; their semantics are defined by the capability schemes themselves and not by this document.

Implementation Status Per , this specification is supported by a reference implementation. The implementation is written in Go (standard library only, no CGO) and covers certificate issuance with authorizationConstraints, AIC extension parsing, DelegationAuthorization signature verification, permission intersection (P (AND) C (AND) T) decisions, capability plugin routing, revocation, and fully offline constraint validation. The companion AIC-JWT profile is implemented in Go and in TypeScript/WebCrypto for browser-compatible runtimes, with a publicly accessible browser demo at https://varwof.org/aicjwt/. The implementation is publicly available at: https://github.com/varwof/types -- shared types, including the AIC extension structures and the AIC-JWT core (Apache-2.0), release v0.3.1; https://github.com/varwof/core -- PKI engine and certificate issuance (AGPL-3.0), release v0.2.0; https://github.com/varwof/engine -- in-memory PKI data engine (AGPL-3.0), release v0.2.0; https://github.com/varwof/gateway-core and https://github.com/varwof/gateway -- zero-trust gateways (Apache-2.0), gateway-core release v0.2.0 and gateway release v0.1.0; https://github.com/varwof/aic-jwt -- AIC-JWT implementation (Go wrapper and TypeScript/WebCrypto pipeline, Apache-2.0); https://github.com/varwof/client -- command-line client (Apache-2.0), release v0.1.0. Language ports of the AIC extension structures and validation logic are also published: C/OpenSSL (https://github.com/varwof/openaic), Java (https://github.com/varwof/aic-lib-java), and .NET (https://github.com/varwof/aic-lib-dotnet), all Apache-2.0. Test status (2026-08-27): all modules build (11/11) and all unit test suites pass (go test ./...), including the certificate authority and HTTP serving packages; the AIC-JWT TypeScript suite passes (node --test). The implementation status of the AIC-JWT profile, including its test suites, is documented in . Interoperability: an independent implementation of the AIC mapping boundary -- the EMILIA crossing profile, which consumes native AIC verifier results and binds them to one exact action and one relying-party admission domain -- reproduces 13/13 positive and hostile conformance cases against the same pinned sources. Publication of the joint conformance kit is in progress. Performance characteristics (informative; measured 2026-08-27 on an 18-core x86 server, engine mode with MySQL; methodology and raw results at https://github.com/varwof/core/docs/bench/en/benchmark-report-2026-08-27.md): Enterprise steady state: 833 AIC certificates/s (500,000 agents, one certificate per agent per 10 minutes) at p50 = 2.4 ms and p99 = 5.0 ms, 0.00% error rate, approximately 7x headroom below the measured ceiling; Single-machine ceiling: approximately 6,100 AIC certificates/s (8,040/s with turbo boost), dominated by ECDSA signing and DelegationAuthorization verification; Worst-case burst: 500,000 simultaneous issuance requests drain in approximately 82 s with no dropped requests when the in-memory engine budget is sized accordingly; Low-end hardware: a Raspberry Pi 5 (4-core, 2.4 GHz) sustains the 833/s steady state at p99 = 330 ms with approximately 2.5x headroom.

Interoperability

Relationship to Standard X.509 Extensions

Key Usage and Extended Key Usage AIC-enabled certificates SHOULD include the digitalSignature key usage ( Section 4.2.1.3). Extended Key Usage MUST include id-kp-clientAuth when the certificate is used for TLS client authentication.

Basic Constraints AIC Agent certificates MUST have BasicConstraints CA:FALSE. CA issuance MUST reject any request where the requester holds an AIC certificate, preventing Agent-to-Agent chaining in single-level deployment mode. In a delegation chain of depth 1, the sub-agent's DelegationAuthorization is anchored to the delegating agent's AIC certificate; as a best practice, the CA SHOULD NOT issue a certificate with DelegationDepthControl.maxDepth greater than 1.

Interoperability Test Matrix Scenario AIC Gateway Behavior Legacy Client Behavior Compatibility Client without AIC extension Reject if RequireAIC=true; pass-through if false Normal mTLS Config-dependent Client with AIC Full admission pipeline AIC extension ignored Yes Multiple capabilities (<=64) Full parse and evaluate Ignored Yes Multiple capabilities (>256) Reject (DoS protection) Ignored Limited Unknown capability scheme Deny Ignored Limited OCSP MUST-Staple missing Reject Normal OCSP No Offline + no CRL cache Fail-Close (reject) Fail-Open (allow) No

TLS Integration

Handshake Requirements The AIC extension does not introduce a new TLS handshake protocol. Instead, it leverages the existing X.509 certificate chain presented during TLS 1.3 handshakes. When a TLS server requires AIC-authenticated connections, it SHOULD include the AIC OID in the CertificateRequest message.

TLS Alert Codes The following is a recommended mapping from AIC conditions to the official TLS alert codes defined in the IANA TLS Alert Registry (); implementations MAY choose appropriate alerts in accordance with the TLS specification. Condition TLS Alert Code Description AIC required but missing unsupported_extension 116 Client MUST include AIC AIC malformed bad_certificate 42 Parsing failure Capability not authorized certificate_unknown 46 Permission denied Impersonation without DA access_denied 49 Missing authorization Certificate expired/revoked certificate_expired 45 Lifecycle check failed

Limitations This specification has the following limitations: Delegation depth limited to 1: Single-level delegation (Principal -> Agent) is the default; a chain of depth 1 (Principal -> Agent -> sub-Agent) MAY be supported when needed. Chains deeper than 1 are not recommended (best practice), because deeper chains complicate attribution and accountability, which this specification anchors to the natural person at the top of the chain. No distributed state: Multi-gateway deployments require out-of-band state synchronization (nonce cache, concurrency tracking). Static capability evaluation: The pure container design defers all semantic validation to gateway plugins. Dynamic, context-aware policy evaluation is not in scope. UDP/DTLS transport: AIC for TCP/TLS and QUIC is fully specified. Non-TLS UDP/DTLS transport is reserved for a future revision. Post-quantum readiness: Algorithm OIDs for hybrid and PQC suites are reserved but signature exchange is not yet specified. Deployment scale: Enterprise-scale validation with >1M agents is not yet published.

Intellectual Property This document is subject to BCP 79 (RFC 8179). The author has filed patent applications related to the technologies described in this document, including Chinese patent applications CN2026112384541 and CN2026112384607 (filed with the China National Intellectual Property Administration). The author has filed IPR disclosure 7553 with the IETF in accordance with BCP 79 requirements. Any applicable IPR disclosures are available through the IETF IPR disclosure system.

Acknowledgments The author thanks the IETF community for ongoing discussions on agent identity and accountability frameworks.

Change Log draft-wei-aic-identity-cert-01 (2026-08-30): Added public reference-implementation repository URLs with pinned release snapshots, test status, independent-implementation interoperability results, and measured performance characteristics (informative) to the Implementation Status section. Added informative reference to the AIC-JWT companion profile . Listed additional language ports (C/OpenSSL, Java, .NET). Aligned patent application numbers with IPR disclosure 7553 (CN2026112384541, CN2026112384607). draft-wei-aic-identity-cert-00 (2026-08-18): * Initial individual draft.