Release 1.21

cert-manager v1.21 includes:

• Removal of the default tokenrequest RBAC from the Helm chart (breaking change)
• Removal of Challenge and Order write permissions from the cert-manager-edit aggregate ClusterRole (breaking change)
• Removal of configurable metrics path and port name Helm values (breaking change)

Major Themes

Default tokenrequest RBAC removed from Helm chart

⚠️ Breaking change

The Helm chart no longer creates a default Role and RoleBinding granting the cert-manager controller permission to create tokens for its own ServiceAccount (serviceaccounts/token: create).

This RBAC was added in v1.16 (cert-manager/cert-manager#7213) to support a "Using the cert-manager ServiceAccount" section in the Route53 documentation. That docs section was subsequently removed (cert-manager/website#1555) when the Route53 page was restructured, and no documented workflow — Route53 IRSA ambient, Vault Kubernetes auth, or any other issuer — requires the controller to mint tokens for its own ServiceAccount.

If you use serviceAccountRef.name pointing at the controller ServiceAccount, you must now either:

• create your own Role and RoleBinding granting serviceaccounts/token: create on that ServiceAccount, or
• migrate to a dedicated ServiceAccount with its own RBAC (recommended — see the Vault or Route53 documentation).

Credit to @everping and @kodareef5 for independently identifying (via privately reported security advisories) that this default RBAC widened the trust boundary beyond what cert-manager's published threat model documents.

Restrict Challenge and Order RBAC in cert-manager-edit ClusterRole

⚠️ Potentially breaking change

The cert-manager-edit aggregate ClusterRole no longer grants create for challenges.acme.cert-manager.io or create, patch, update for orders.acme.cert-manager.io. This fixes a security issue (GHSA-8rvj-mm4h-c258) where these permissions allowed namespace users to bypass Issuer solver selectors and abuse ClusterIssuer credentials.

These resources are internal to cert-manager's ACME workflow and are not intended to be created or modified directly by users. Challenge patch and update are retained because the Challenge spec is immutable after creation and users may need these verbs to remove stuck finalizers (cert-manager/cert-manager#3851, cert-manager/cert-manager#3870).

This change was already shipped in patch releases v1.20.3 and v1.19.6, so if you are already running one of those versions this will not be a breaking change.

If you have tooling or workflows that create Challenge or Order resources directly (outside of the normal Certificate → CertificateRequest → Order → Challenge flow), you will need to grant those permissions explicitly.

Metrics port name and path Helm values removed

⚠️ Breaking change

The Helm values prometheus.servicemonitor.targetPort, prometheus.servicemonitor.path, and prometheus.podmonitor.path have been removed. The metrics path is always /metrics and the ServiceMonitor targetPort is always the port named http-metrics — these were the defaults and there was no supported reason to override them.

The controller Service metrics port has also been renamed from tcp-prometheus-servicemonitor to http-metrics, aligning it with the webhook and cainjector services.

Because the Helm values schema uses additionalProperties: false, users who still have any of the removed keys in their values overrides will see a schema validation error on upgrade. Remove these keys from your values file before upgrading.

(cert-manager/cert-manager#8952, @erikgb)

Configurable CertificateRequest retry backoff duration

cert-manager 1.21 adds the --certificate-request-maximum-backoff-duration controller flag (default: 32 hours), making the exponential backoff cap configurable alongside the existing --certificate-request-minimum-backoff-duration flag.

When a CertificateRequest fails, cert-manager backs off exponentially — by default from 1 hour up to 32 hours. In environments with scheduled CA maintenance windows, the backoff can grow so large that cert-manager does not retry until hours after the CA comes back online. With this flag, cluster operators can lower the ceiling to match their CA's maintenance schedule. For example, --certificate-request-maximum-backoff-duration=1h ensures that cert-manager retries at most every hour regardless of how many consecutive failures have occurred.

The default of 32 hours preserves the behavior of previous releases. The flag may also be set via the certificateRequestMaximumBackoffDuration field in the controller ControllerConfiguration API, or via the Helm chart:

config:
  certificateRequestMaximumBackoffDuration: 1h

See the controller CLI reference for the full list of flags, and What happens if issuance fails? Will it be retried? for background on how cert-manager retries failed issuances.

(cert-manager/cert-manager#8893, @lunarwhite)

Skip the self-check with waitInsteadOfSelfCheck

cert-manager 1.21 adds the waitInsteadOfSelfCheck solver option for ACME HTTP01 and DNS01 challenges. When set, cert-manager skips its own self-check and instead waits the configured duration after presentation before asking the ACME server to validate. This is an escape hatch for environments where cert-manager cannot reliably observe the same validation path as the ACME server, such as split-horizon DNS or NAT loopback (or hairpinning).

See Skip the self-check with waitInsteadOfSelfCheck for configuration details.

Webhook Serving Certificate Renewal After System Suspend

The cert-manager webhook generates and automatically renews its own self-signed serving certificate. Prior to 1.21, the renewal timer relied solely on Go's monotonic clock (CLOCK_MONOTONIC). During system suspend (S3/S4) or VM live migration, CLOCK_MONOTONIC stops advancing while wall-clock time (CLOCK_REALTIME) continues. When the system resumes, the renewal timer has not yet reached its deadline, so the webhook's serving certificate is never renewed — even though it has already expired. This causes x509: certificate has expired or is not yet valid errors for all admission and conversion webhook calls.

cert-manager 1.21 adds a periodic ticker that polls wall-clock time against the renewal deadline, detecting missed renewals regardless of whether CLOCK_MONOTONIC advanced. The webhook now recovers within one minute of system resume.

More details are available in the PR: https://github.com/cert-manager/cert-manager/pull/8464.

Community

As always, we'd like to thank all of the community members who helped in this release cycle, including all below who merged a PR and anyone that helped by commenting on issues, testing, or getting involved in cert-manager meetings. We're lucky to have you involved.

A special thanks to:

• @Copilot
• @FelixPhipps
• @Peac36
• @SebTardif
• @apkatsikas
• @bitloi
• @dap0am
• @figaw
• @immanuwell
• @jabbrwcky
• @jnohlgard
• @jsoref
• @ltwongaa
• @lunarwhite
• @mateenali66
• @onurmicoogullari
• @putongyong
• @seanorama
• @texasich

for their contributions, comments and support!

Also, thanks to the cert-manager maintainer team for their help in this release:

• @SgtCoDFish
• @ThatsMrTalbot
• @erikgb
• @hjoshi123
• @inteon
• @maelvls
• @munnerz
• @wallrj
• @wallrj-cyberark

And finally, thanks to the cert-manager steering committee for their feedback in this release cycle:

• @FlorianLiebhart
• @TrilokGeer
• @ianarsenault
• @ssyno

v1.21.0

Changes since v1.20.0:

Feature

• Add Venafi OAuth token request observability and a new AuthFailed Issuer condition reason to distinguish bad credentials from transient infrastructure errors. (#8808, @FelixPhipps)
• Add certificateRequestMaximumBackoffDuration controller configuration option to cap retry backoff time for failed CertificateRequests. Configurable via config file, --certificate-request-maximum-backoff-duration CLI flag, or Helm value config.certificateRequestMaximumBackoffDuration. Defaults to 32 hours for backward compatibility. (#8893, @lunarwhite)
• Add an optional waitInsteadOfSelfCheck field to ACME HTTP01 and DNS01 solvers so cert-manager can skip its own self-check and ask the ACME server to validate after a configured wait. (#8858, @wallrj)
• Add configurable runtimeClassName support for cert-manager components and ACME HTTP01 solver pods. (#8791, @jsoref)
• Add new controller flag --acme-http01-solver-extra-labels, allowing Helm's global.commonLabels to propagate to all dynamically-created ACME HTTP01 solver resources (Pods, Services, Ingresses, or Gateway API HTTPRoutes). (#8761, @lunarwhite)
• Add opt-in startupapicheck.ttlSecondsAfterFinished Helm value to enable automatic cleanup of the startupapicheck Job via the Kubernetes TTL-after-finished controller. (#8523, @dap0am)
• Added ARI support through the ACMEUseARI feature gate. (#8798, @hjoshi123)
• Added AWS IAM authentication support for Vault issuer, including IRSA (IAM Roles for Service Accounts) and ambient credentials (EC2/ECS). (#8422, @bitloi)
• Added cert-manager.io/ignore-tls-listeners annotation for ignoring gwapi listeners. (#8727, @hjoshi123)
• Added option to specify additional listener protocols the GatewayAPI integration will consider when creating certificates. (#8683, @ThatsMrTalbot)
• Adds support for the Modern2026 go-pkcs12 profile and FIPS 140-3 (#8841, @seanorama)
• Cainjector:
  • A new flag --ignore-namespaces was added to the cainjector binary. It can be used to filter out namespaces from being watched for secrets to use for injectables. (#8614, @figaw)
• Disabled client side rate-limiting if AP&F is enabled. (#8757, @hjoshi123)
• Extend the Venafi/CyberArk integration to also support PANW NGTS. (#8779, @FelixPhipps)
• Feat(certificate): adding certificate renewal policies (#8258, @hjoshi123)
• Make cainjector use SSA unconditionally and deprecate the ServerSideApply feature gate (#8692, @erikgb)
• Processed annotations cert-manager.io/alt-names, cert-manager.io/ip-sans to Certificates generated from ingress like objects in cert-shim controllers. (#8927, @jabbrwcky)
• Promote the CAInjectorMerging feature gate to GA (#8583, @Copilot)
• When using ACME HTTP-01 with a ListenerSet, setting the annotation acme.cert-manager.io/http01-parentreffallback: "true" causes cert-manager to use the parent Gateway as the solver HTTPRoute parentRef instead of the ListenerSet. This enables TLS-only ListenerSets to rely on a shared Gateway HTTP listener for ACME challenges. (#8749, @apkatsikas)

Bug or Regression

• BREAKING: The Helm chart no longer ships a default Role and RoleBinding granting the cert-manager controller ServiceAccount permission to create tokens for itself (serviceaccounts/token: create). This RBAC was added in v1.16 (#7213) but no documented workflow requires it, and the motivating Route53 docs section was removed in Oct 2024. If you rely on serviceAccountRef.name pointing at the controller ServiceAccount (an undocumented pattern), you must now create your own Role and RoleBinding granting serviceaccounts/token: create on that ServiceAccount, or migrate to one of the documented patterns (IRSA ambient, or a dedicated ServiceAccount with its own RBAC). (#8931, @wallrj-cyberark)
• ACME challenges no longer terminally fail on transient network errors (TLS handshake timeouts, DNS failures, context cancellation) during nonce fetches and authorization waits. The challenge controller returns the error and lets the workqueue retry with backoff. (#8760, @texasich)
• Add dns issuer secrets validation before marking it as ready (#8255, @Peac36)
• Add missing issuer finalizer RBAC to the order controller to support owner references (#8654, @erikgb)
• ClusterIssuer metrics collector now correctly respects the enabled-controllers configuration, avoiding a redundant startup when only operating within a namespace. (#8822, @lunarwhite)
• Fix Venafi TPP issuer setup and signing regression on master: restore authentication of the vcert connector in the client constructor, which was removed in #8808. (#8843, @wallrj-cyberark)
• Fix a performance issue in the certificateRequestApproval webhook where CertificateRequests referencing a GroupKind whose CRD is not yet installed would trigger repeated API server discovery queries on every admission request. Negative results are now cached for 30 seconds. (#8651, @mateenali66)
• Fix webhook serving certificate not being renewed after system suspend. (#8464, @Peac36)
• Fixed a rare panic in the trigger controller when a Certificate is deleted from the informer cache while a reconcile is in progress (e.g. during namespace teardown). (#8962, @hjoshi123)
• Fixed an integer overflow in renewBeforePercentage calculations that caused Certificates with durations longer than approximately 3 years to be incorrectly rejected by validation or assigned incorrect renewal times. (#8947, @ThatsMrTalbot)
• Fixed duplicate parentRef bug when both issuer config and annotations are present. (#8619, @hjoshi123)
• Fixed infinite re-issuance loop when issuer returns an already expired certificate (#8610, @onurmicoogullari)
• Fixed local e2e-setup-samplewebhook installation to use the samplewebhook image repository and tag from the saved image tarball manifest. (#8821, @wallrj)
• Fixed potential OOM in DNS-over-HTTPS client by bounding response body read with io.LimitReader (128 KB cap). (#8803, @SebTardif)
• Fixed validation of timezone-prefixed renewal window cron specs without a schedule. (#8813, @immanuwell)
• Harden ACME Challenge and Order resources: reject user-created Challenges without Order ownership, enforce Order spec immutability, and detect pre-placed same-name Challenges with mismatched specs. (#8948, @wallrj-cyberark)
• Helm chart bugfix: rename image helper to avoid umbrella chart conflicts (#8753, @FelixPhipps)
• Helm: Fix invalid YAML generated when both webhook.config and webhook.volumes are defined. (#8664, @jnohlgard)
• Remove ACME Challenge create and Order create/patch/update from the cert-manager-edit aggregate ClusterRole to prevent direct manipulation of these internal resources (GHSA-8rvj-mm4h-c258). (#8958, @wallrj-cyberark)
• Remove issuer owner reference from challenges blocking challenge garbage collection (#8743, @erikgb)
• Update logic to identify and preserve the secret matching nextPrivateKeySecretName (#8577, @putongyong)
• Vault Issuer webhook validation now rejects .. path segments in spec.vault.path and auth mount path fields, preventing path.Join from silently resolving relative segments before constructing the Vault API request. (#8930, @wallrj-cyberark)

Other (Cleanup or Flake)

• API cleanup: removed deprecated ObjectReference (#8625, @inteon)
• Remove Helm values prometheus.servicemonitor.targetPort, prometheus.servicemonitor.path, and prometheus.podmonitor.path. The metrics path is always /metrics and the target port is always http-metrics. Rename the controller service metrics port from tcp-prometheus-servicemonitor to http-metrics for consistency with other workloads. Users must remove these keys from their value overrides before upgrading. (#8952, @erikgb)
• The enableGatewayAPI and enableGatewayAPIListenerSet fields on ControllerConfiguration are deprecated and moved into the gatewayAPI sub-struct as gatewayAPI.enabled and gatewayAPI.enableListenerSet. The old fields continue to work. (#8732, @ThatsMrTalbot)
• Update base images to Debian 13 (#8849, @ltwongaa)