# What is X.509?

> X.509 is the standard format for a public-key certificate: a record, signed by a certification authority, that ties a public key to a named person, server or workload for a limited time.

Explainer by Boston Identity, 2026-10-28. 6 min read. https://insights.bostonidentity.com/glossary/x509

- What it is: standard
- Defined by: ITU-T X.509; IETF profile RFC 5280
- Who it's for: Teams issuing identities to servers, workloads and agents

## What is X.509?

A public key is only useful if you know whose it is. X.509 answers that with a certificate: a data structure that binds a public key to a subject, the person or system that holds the matching private key [S1 §3.1]. A certification authority (CA) asserts the binding by signing the certificate [S1 §3.1]. Because anyone can check the signature and the dates, certificates can travel over untrusted networks and sit in unsecured storage [S1 §3.1].

The format comes from the International Telecommunication Union (ITU-T), which first published it in 1988; version 3 added extension fields and was completed in June 1996 [S1 §3.1]. RFC 5280, from the Internet Engineering Task Force (IETF), is the profile that says how Internet software should use version 3 certificates and certificate revocation lists (CRLs) [S1 §1]. It names web, email, user authentication and IPsec among the uses [S1 §2].

Every certificate carries the subject's and issuer's names, the subject's public key, a validity period, a version and a serial number [S1 §4.1.2]. Most also carry extensions, such as the subject alternative name, which binds a DNS name, IP address, email address or URI to the key [S1 §4.2.1.6].

Think of it like this: A certificate works like a passport: an issuer everyone recognises vouches that this face belongs to this name, until a printed expiry date. The comparison stops at revocation, because a border guard can phone the passport office, while a certificate user may only learn of a cancellation when the next published list comes out.

Not the same as: X.509 is not a protocol for logging in or connecting. It is the certificate format; TLS, mutual TLS and SPIFFE are among the systems that carry and check X.509 certificates.

## How does X.509 work?

Four parties take part: the subject that holds a key pair, the CA that signs, a repository that stores certificates and CRLs, and the relying party that wants to trust the subject's key [S1 §3].

[Figure: The CA's signature binds the subject's name to its public key, and the relying party trusts that binding only after chaining it back to a CA key it was loaded with and checking the latest revocation list.]

1. **The subject registers with a CA.** The subject first makes itself known to a CA, either directly or through a registration authority (RA), a system the CA hands some management work to [S1 §3, §3.5]. The subject's client typically also needs its own key pair [S1 §3.5].
2. **Clients are loaded with trusted CA keys.** Before it can check anything, a client must be loaded with the public key of at least one CA it trusts [S1 §3.5]. A path starts from a CA in the user's own domain or from the top of a hierarchy, and a self-signed certificate is how such a starting key is conveyed [S1 §3.2].
3. **The CA signs the certificate.** The CA fills in the subject's name, public key, validity dates, a serial number unique to that CA, and any extensions, then signs the whole block [S1 §4.1.2, §4.1.2.2, §4.1.1.3]. By signing, the CA certifies the binding between the key and the subject [S1 §4.1.1.3]. Every part of a subject alternative name must be verified by the CA [S1 §4.2.1.6].
4. **The certificate is published.** The CA returns the certificate to the subject's system, posts it in a repository, or both [S1 §3.5]. Any means of distribution may be used; a directory is not required [S1 §2.1].
5. **The relying party builds a path.** A relying party usually needs a chain: the subject's certificate signed by one CA, plus zero or more certificates of CAs signed by other CAs, back to a key it already trusts [S1 §3.2]. It links each certificate to the next by matching the issuer name in one with the subject name in a CA certificate [S1 §4.1.2.4]. A key may sign certificates only if its certificate says it belongs to a CA [S1 §4.2.1.9].
6. **The relying party checks signature, dates and status.** The relying party checks the certificate's signature and validity period, then fetches a suitably recent CRL and confirms the serial number is not on it [S1 §3.3]. It must reject a certificate with a critical extension it does not recognise [S1 §4.2]. Online checking, such as the Online Certificate Status Protocol (OCSP), may stand in for CRLs [S1 §3, §3.3].

## Why does X.509 matter now?

Guidance on agent security now names certificates directly. Anthropic's agent security guide asks for a cryptographic identifier for every agent instance, and its higher tiers add X.509 certificates with rotation and revocation, then mutual TLS [BI-FIND-2026-0596](https://insights.bostonidentity.com/ref/BI-FIND-2026-0596). It also says static API keys are not acceptable even at its entry tier, and points to certificates or short-lived tokens bound to the agent [BI-FIND-2026-0542](https://insights.bostonidentity.com/ref/BI-FIND-2026-0542).

The IETF's workload identity drafts lean on the same format. The Workload Identity Certificate is an ordinary X.509 certificate with the workload's identifier in one URI subject alternative name, and its implementation section calls it compatible with the certificates SPIFFE deployments already issue [BI-FIND-2026-0375](https://insights.bostonidentity.com/ref/BI-FIND-2026-0375). An agent-authentication draft lists an X.509 credential among its accepted formats and leaves the choice to each deployment [BI-FIND-2026-0179](https://insights.bostonidentity.com/ref/BI-FIND-2026-0179).

New drafts also try to load more into the certificate itself. One profile writes an agent's owner, trust level and capabilities into its certificate, valid for up to 365 days [BI-FIND-2026-0748](https://insights.bostonidentity.com/ref/BI-FIND-2026-0748). Another draft argues the opposite, that capabilities and operator must never sit in the identifier [BI-FIND-2026-0754](https://insights.bostonidentity.com/ref/BI-FIND-2026-0754). Which model a deployment picks decides whether a change of permissions means reissuing the agent's identity.

## Where does it fit?

- Compared with SPIFFE: SPIFFE is a framework for handing identities to workloads; one of the forms it issues is an X.509 certificate with a spiffe URI as the subject [BI-FIND-2026-0375](https://insights.bostonidentity.com/ref/BI-FIND-2026-0375). X.509 is the format, SPIFFE decides how workloads get one.
- Compared with mTLS: Mutual TLS is the connection in which both sides present a certificate. The IETF's workload mutual TLS draft requires exactly one workload identifier URI subject alternative name, a path to a trust anchor configured for the trust domain, and a revocation check under the PKIX rules [BI-FIND-2026-0517](https://insights.bostonidentity.com/ref/BI-FIND-2026-0517).
- Compared with JWT: One workload metadata draft carries attributes as claims inside a JSON Web Token (JWT) credential, but gives an X.509 certificate only an extension that points to a metadata endpoint [BI-FIND-2026-1355](https://insights.bostonidentity.com/ref/BI-FIND-2026-1355).
- Compared with Token Status List: Token Status List is a revocation method for tokens, which the SD-JWT standard names only as an example [BI-FIND-2025-0145](https://insights.bostonidentity.com/ref/BI-FIND-2025-0145). X.509 revokes certificates through a CA-signed CRL that names revoked certificates by serial number, or through an online service [S1 §3.3].

## What goes wrong?

- The CA's basis for signing can be thin. A CA may base the binding on a proof of possession, on presentation of the private key, or on the subject's own assertion [S1 §3.1].
- Revocation lags. A CRL is reissued on a schedule, such as hourly, daily or weekly, so a revocation may not reach relying parties until the next one [S1 §3.3]. Online checking shortens the delay, but the relying party then has to trust the online service [S1 §3.3].
- Unknown extensions can be skipped. A relying party may ignore a non-critical extension it does not recognise [S1 §4.2]. A draft agent profile marks its authority extension non-critical, so software that does not know it accepts the certificate and skips every limit in it [BI-FIND-2026-0748](https://insights.bostonidentity.com/ref/BI-FIND-2026-0748).
- The CA's checks set the ceiling. The same profile says the issuing CA is not required to check whether the requested capabilities make sense [BI-FIND-2026-0748](https://insights.bostonidentity.com/ref/BI-FIND-2026-0748). Another draft scores agents at an issuing CA that each operator typically runs for itself, and where neither the attestation nor the score comes from outside, we read that as operators grading their own agents [BI-FIND-2026-1291](https://insights.bostonidentity.com/ref/BI-FIND-2026-1291).
- Attributes may belong elsewhere. RFC 5280 itself says that as tools such as attribute certificates emerge, it may be appropriate to limit the authenticated attributes in a certificate [S1 §2].
- Certificates can leak what they contain. An ACME draft for device certificates warns that a hardware identifier in a certificate logged to certificate transparency becomes permanently public [BI-FIND-2026-1052](https://insights.bostonidentity.com/ref/BI-FIND-2026-1052).
- Long lives need a plan for their end. When an issuer cannot keep status information until the expiry date, it must make sure no valid path to the certificate remains [S1 §4.1.2.5].

## What should a security leader ask?

1. Which CAs do our servers, workloads and agents trust as starting points, and who can add one?
2. How long do our workload and agent certificates live, and how fast do they rotate?
3. How do our relying parties check revocation, and how stale can a CRL or status answer be before they reject a certificate?
4. Which non-critical extensions do we rely on for access decisions, and does every service that checks the certificate understand them?
5. What does our issuing CA verify before it signs, and does anyone outside the team that runs the agents check it?

## Sources

- [S1] [Internet X.509 Public Key Infrastructure Certificate and Certificate Revocation List (CRL) Profile](https://www.rfc-editor.org/rfc/rfc5280) (IETF, RFC 5280, Standards Track, May 2008)

## Evidence

- [BI-FIND-2026-0596](https://insights.bostonidentity.com/ref/BI-FIND-2026-0596): Anthropic requires a cryptographic identifier for each agent instance at its Foundation tier. "Every agent instance needs a unique, cryptographically rooted identifier that persists across its actions", and "Unique identifiers alone are a labeling exercise". Service tokens should be measured in minutes, and static API keys and shared service-account passwords "are no longer a legitimate entry point, not even at Foundation". Higher tiers add X.509 certificates with rotation and revocation, mutual TLS, then hardware-backed identity with remote attestation, and access control climbs from deny-by-default roles to per-action authorization with immediate revocation. The standards named are OAuth 2.0, X.509, mTLS, FIDO2 and OpenTelemetry; SPIFFE, WIMSE and token exchange do not appear. (from "Anthropic asks for per-agent identity; its guide leaves Claude Code sub-agent credentials unspecified")
- [BI-FIND-2026-0542](https://insights.bostonidentity.com/ref/BI-FIND-2026-0542): Anthropic's guide raises the entry bar for an AI agent. Its framework has three cumulative tiers, Foundation, Enterprise and Advanced, with Foundation meant for small businesses and teams and Enterprise "where most organizations should aim". Even Foundation requires a persistent agent identity "backed by cryptographic material (not just labels)", OAuth 2.0 or similar tokens "with expiration measured in minutes", and no static API keys, embedded credentials or shared service-account passwords: "they are no longer a legitimate entry point", and existing ones should be treated as already compromised. "Static API keys are not acceptable for tool authentication, even at Foundation"; the guide's suggested route is certificates or short-lived tokens bound to the calling agent. The stated test for any control is whether it makes an attack impossible or only tedious, because AI-accelerated attackers will wait out friction. (from "Anthropic says agents should never use static API keys, yet its sub-agents can carry the parent's permissions")
- [BI-FIND-2026-0375](https://insights.bostonidentity.com/ref/BI-FIND-2026-0375): Half of this draft is already running and the other half is not, and the draft says so. The Workload Identity Certificate is an ordinary X.509 certificate with the workload identifier in exactly one URI SubjectAltName, which the implementation section calls 'fully compatible with the X509-SVID' that SPIFFE deployments have issued for years; SPIFFE deployments simply use their spiffe URIs as the subject. The new piece is the JWT form, and the self-reported status reads beta at SPIFFE, alpha at Defakto Security and research at Teleport, against production for the certificate at all three. The IETF is standardizing the certificate the industry already has and attaching a token nobody has shipped yet. (from "The IETF's workload credential binds one key to one name and says nothing else about the software holding it")
- [BI-FIND-2026-0179](https://insights.bostonidentity.com/ref/BI-FIND-2026-0179): The draft's sharpest rule reads "The Large Language Model MUST NOT have access to an agent's credentials or to credentials that may be needed to access tools and services", which it says keeps prompt injection from getting the model to use, expose or disclose them. Around it sit the identity rules. An agent MUST be assigned exactly one WIMSE identifier and MUST hold credentials cryptographically bound to it; those credentials SHOULD be short-lived and MUST carry an explicit expiry, and static API keys are "an antipattern for agent identity". The draft names WIT, X.509 WIC and SPIFFE SVIDs as formats and leaves the choice to the deployment. Our briefing on Sierra's cloud agent platform found a similar boundary built in: a proxy approves each call and swaps in the credential, so the agent never holds it. (from "Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reach")
- [BI-FIND-2026-0748](https://insights.bostonidentity.com/ref/BI-FIND-2026-0748): Sharif's profile writes an agent's authority into its identity certificate: the owning organization, a trust level from L0 to L4 defined by the actions it allows (L3 may pay and transfer; L4 may deploy and delete, with human approval), a capability list, a delegation depth and a kill-switch URL, in a certificate that may last up to 365 days. The issuing CA checks that the request came from an authorized platform and that the organization controls its SPIFFE domain, but it is 'NOT required to validate the semantic correctness of the requested capabilities'. The extension must be marked non-critical, so any X.509 stack that does not know it accepts the certificate and skips every limit. (from "Sharif's profile writes an AI agent's rights to pay and delete into an identity certificate valid up to a year")
- [BI-FIND-2026-0754](https://insights.bostonidentity.com/ref/BI-FIND-2026-0754): Two agent-identity drafts posted a week apart take opposite positions on what an identity should hold. Sharif's X.509 profile puts the owning organization, trust level and capabilities inside the agent's certificate; Drake's requirement R5 says trust tier, role, operator and capabilities must never be encoded into the identifier and belong in separate, changeable authorizations. Whichever model a deployment takes decides whether a change of operator or permissions means reissuing the agent's identity. (from "Drake's case for an agent identity that outlives its keys and its operator, made without weighing OAuth, SPIFFE or X.509")
- [BI-FIND-2026-0517](https://insights.bostonidentity.com/ref/BI-FIND-2026-0517): WIMSE mutual-TLS revision -03 requires a peer certificate to validate to a trust anchor configured for the trust domain in its workload identifier, and requires certificate revocation status to be determined under the referenced PKIX rules. The certificate must contain exactly one workload-identifier URI subject alternative name; zero or multiple such entries must be rejected. On a DNS-based connection, the client must also reject a missing or mismatching DNS identity even when workload-identifier validation succeeds. Authorization must retain the trust domain and the trust anchor rather than evaluate only an identifier path. (from "WIMSE mutual TLS -03 ties certificate acceptance to trust-domain anchors")
- [BI-FIND-2026-1355](https://insights.bostonidentity.com/ref/BI-FIND-2026-1355): Metadata travels in one of two ways: as claims inside a JWT credential, or from an authenticated metadata endpoint. The draft defines no way to carry the attributes themselves in an X.509 certificate, only a non-critical extension pointing to that endpoint, and JWT credentials can carry the pointer as an optional metadata_uri claim. An endpoint configured out of band takes precedence over the pointer. When neither exists, a client may use a default /.well-known/workload-metadata address if its policy trusts that domain. The endpoint authenticates each caller and discloses claim by claim, so knowing a workload's identifier does not get you its metadata, and an omitted claim asserts nothing. A selected endpoint that fails MUST NOT trigger fallback to a lower-precedence one. If required metadata cannot be had from an acceptable source, access that depends on it MUST NOT be granted. A response is not a signed assertion meant for forwarding to another party; signed responses are one of the draft's open issues. (from "An individual IETF draft takes a workload's role and group out of its name and puts them in separate metadata")
- [BI-FIND-2025-0145](https://insights.bostonidentity.com/ref/BI-FIND-2025-0145): What RFC 9901 hands to the protocol or profile on top is, as we read it, most of what an agent deployment has to settle: how the holder's key pair is established (section 4.1.2), how the key-binding JWT's audience is represented and how its nonce is obtained (4.3), if and how consent to disclose is obtained (7.2), which validity claims a verifier must see, which application-specific profiles SHOULD list (9.7), and how verifiers get issuers' keys, which it does not define (9.8). It defines no status or revocation check for an issued SD-JWT; Token Status List appears only as an example of a revocation callback that tells the issuer little. In our briefings these gaps mostly carry forward: Verifiable Intent v0.1 and the Delegate SD-JWT draft define no revocation, and the SD-Card draft puts the registry that issues agent cards out of scope. Our briefing on SD-JWT VC found that it answers the issuer-key question with a well-known metadata document or an X.509 chain. (from "The SD-JWT standard that agent drafts build on never mentions agents and leaves key binding to each verifier")
- [BI-FIND-2026-1291](https://insights.bostonidentity.com/ref/BI-FIND-2026-1291): APKI writes a trust score from 0 to 100 into each agent's certificate, and relying parties MUST reject a certificate scored below the minimum tier they set. The score comes from a Trust Score Computation Engine at the issuing CA, which the draft says is typically the intermediate CA that each organization running agents deploys for itself. The starting score adds fixed allowances (up to 30 points for attestation strength, 20 for organization reputation, 15 each for model provenance and code attestation, 20 for history), and relying parties' reports then move it by set ranges. The draft gives no basis for any of these numbers and leaves the reporting protocol to a companion specification. Against score inflation, behavioural signals MUST be authenticated with the reporting relying party's credentials, scores SHOULD draw on several independent sources and "no single relying party should be able to move a score by more than 15 points", and the CA MUST NOT issue a certificate scoring above the agent's historical peak minus any penalty history. Another organization MAY apply its own mapping to a partner's scores; the draft's example counts them 10 points lower. A separate field records how the agent's capabilities were attested, from self-declared through CA-verified and third-party to hardware-bound: relying parties requiring high assurance MUST require CA-verified or higher, and third-party attestation SHOULD be required for regulated uses such as financial services. CA-verified means the issuing CA tested the agent, and the draft describes third-party as an independent auditor's check. The draft answers a malicious operator with transparency logs, third-party attestation and federated trust with minimum score thresholds, and answers an insider with separation of duties between the CA operator and score computation; section 6.3 still places the score engine at the issuing CA. Where neither the attestation nor the score comes from outside, we read the score as the agent's operator grading its own agents. (from "An agent certificate draft ties access to a trust score set by the issuing CA, usually the agent operator's own")
- [BI-FIND-2026-1052](https://insights.bostonidentity.com/ref/BI-FIND-2026-1052): At -01, putting a device's identifiers in the certificate was already optional: a client MAY include the permanent identifier in its certificate request, a CA that wants only privacy-preserving certificates MAY reject requests that carry it, and a CA MAY leave the hardware-module identity out. By -10, posted in August 2026, the authors had turned the option into advice and a change to ACME itself. The draft now updates RFC 8555 so that, for these two identifier types, the certificate request SHOULD rather than MUST repeat the identifiers in the order. It tells implementers to treat leaving them out as the default unless there is an operational need to include them, warns that a hardware identifier in a certificate logged to certificate transparency becomes permanently public, and for workload mutual TLS advises preferring logical identifiers such as SPIFFE URIs, with the hardware attestation used only to bootstrap. (from "An ACME draft lets company CAs issue device certificates from the device's own proof, and each CA judges it")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/x509 as the source. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
