Temporal Technologies, Inc.'s temporal-proxy versions 0.2.0 through 0.8.0 have an authorization bypass vulnerability due to improper binding of encrypted payloads to namespaces. The proxy selects the decryption key based on an identifier in the payload rather than the request's namespace, allowing an authorized caller for one namespace to decrypt payloads from another namespace. This does not expose new encrypted payloads to attackers nor allow modification of ciphertext. The vulnerability has a medium severity score of 6.0 CVSS v4.0.
Temporal Server versions 1.30.0 through 1.30.7, 1.31.0 through 1.31.3, and 1.32.0 are affected by an authorization bypass vulnerability in Scheduler V2. An authenticated user with write permission in one namespace can manipulate schedule state in another namespace by exploiting internal completion callbacks that are not properly bound to namespaces. This can cause schedules to incorrectly record actions as completed or pause schedules improperly. No workflow payload or history disclosure is involved. Scheduler V2 is off by default, and only namespaces with Scheduler V2 schedules are vulnerable.
CVE-2026-104355 is a high-severity authorization vulnerability in Temporal Technologies, Inc.'s Temporal Server affecting multi-cluster deployments. A connected peer cluster can modify local namespaces on another cluster without proper authorization because the receiving cluster applies inbound namespace replication updates without verifying if the namespace is global. This allows a compromised or malicious peer to overwrite namespace metadata such as state, description, retention, and active cluster settings, potentially disrupting workflow operations. Clusters without remote clusters registered with connection enabled are not affected.
All published s2s-proxy versions through 0.2.2 are affected. In versions 0.1.16 through 0.2.2, TLS server listeners use Go's RequireAnyClientCert mode when skipCAVerification is false. This mode checks that the client holds the certificate's private key but does not verify the certificate against the configured CA. An attacker can therefore use a self-signed certificate and key to establish a TLS and yamux connection, then invoke RPCs allowed by the proxy's configuration and Temporal credentials. No certificate or private key trusted by the deployment, and no Temporal credential, is required.
Temporal Server compiles a Worker Controller Instance module into its Worker Service, and that module registers a compute provider named subprocess whose function is to launch a worker by running a command on the machine hosting the Worker Service. The program name and the argument vector that provider executes are taken from the compute provider configuration supplied in the caller's request rather than from operator configuration. An authenticated caller holding only a write role in a single namespace can therefore configure a worker deployment version so that the Worker Service executes a command of the caller's choosing on its own host, under the account the server process runs as. Execution is immediate rather than deferred: the configuration handler invokes every provider using the invoke strategy directly after validating the submitted specification, so no scaling decision, task arrival, or unusual request sequence is required. Because the Worker Service process holds the persistence credentials for every namespace in the cluster and the cluster's TLS material, the consequence reaches beyond the caller's namespace to the cluster as a whole. The provider is present in the official temporal-server binaries and container images for the affected releases. The only control that can keep it unreachable is the compute provider allowlist, the per-namespace dynamic configuration setting workercontroller.compute_providers.enabled, and that control does not deny by default: its default value is an unset list, and the allowlist check is skipped entirely when the value is unset, so every registered compute provider is permitted, this one included. To determine whether a deployment is affected, check the following together. The deployed Temporal Server version is 1.31.0 or later and earlier than 1.31.3. The Worker Service is running, which it is in the default service set and therefore in a stock deployment. The effective per-namespace value of workercontroller.compute_providers.enabled is either unset or contains subprocess. And authorization is configured, meaning a real authorizer and claim mapper are in place; a deployment running with no authorizer already grants every caller unrestricted access to every namespace, so it has no namespace boundary for this to cross. Note that the separate per-namespace dynamic configuration setting workercontroller.enabled does not gate the affected path. It defaults to false, and a deployment that has never set it in any namespace is still affected, which was confirmed by running an affected release with no value for that setting present anywhere in dynamic configuration. To look for a compute configuration that is already attached, call DescribeWorkerDeploymentVersion for each worker deployment version in each namespace and check whether any scaling group's compute provider type is subprocess.
Temporal Server decided whether a Workflow completion callback was internal by reading a caller-supplied HTTP header. An authenticated caller holding only write permission in a single namespace could attach a completion callback whose URL host matched the configured callback address allowlist, whose URL path was any Temporal HTTP API route, and whose header map contained a non-empty header named source. When the History service delivered that callback, the non-empty source header caused it to re-target the request at the local frontend client and rewrite only the scheme and host, preserving the caller's path, query, and request body. Where an internal frontend is deployed with its HTTP API enabled, that client resolves to the internal frontend, which authorizes every request as a system administrator without requiring authentication information. The result is that the server performs an attacker-chosen state-changing HTTP POST against its own administrative API on the caller's behalf, in namespaces where the caller has no permission. The caller never needs network access to the internal frontend, because the History service makes the request. Confirmed effects include terminating Workflows in other namespaces, registering namespaces, modifying another namespace's configuration, and deleting another namespace and its Workflows. The affected routing logic is present in both the HSM and CHASM callback delivery implementations. This description and the CVSS score in this record describe releases 1.30.0 and later, where any non-empty source header is sufficient. Releases 1.25.0 through 1.29.7 are affected by a narrower form of the same defect in which the header must exactly match a configured cluster ID, a UUID that a namespace-scoped caller cannot read through the API. The consequence once that match occurs is the same, but the attack is materially harder and scores lower. To determine whether a deployment is affected, check two settings together: the static Server configuration for a non-zero services.internal-frontend.rpc.httpPort, and the effective per-namespace dynamic configuration value of component.callbacks.allowedAddresses. A deployment is exposed only when an internal frontend is deployed with a non-zero HTTP port, at least one allowlist rule admits a host, and authorization is enabled. The allowlist is empty by default, which denies all external callback URLs, and the stock static topology does not include an internal frontend. Note that a failed attach returns the error 'invalid url: url does not match any configured callback address', which proves only that one tested URL did not match and does not prove the effective allowlist is empty. To look for callbacks already attached, use DescribeWorkflowExecution, which returns callback information for a Workflow's registered completion callbacks. Operators should be aware of a gap when searching for evidence of delivery: the frontend HTTP API server records the request method and URL at debug level only, so at default log levels a delivered request is not written to the internal frontend's logs, and the absence of such log entries is not evidence that the issue was not exploited.
github.com/temporalio/ringpop-go enforces configured LabelOptions limits when an application changes the local node's labels, but affected versions do not apply those limits to label maps received in SWIM membership changes. A network peer that can reach a live Ringpop TChannel listener can repeatedly submit changes for distinct member addresses containing label keys, values, or counts that exceed the receiver's configured limits. Accepted labels are retained in the member list and disseminated to peers, allowing memory and gossip-bandwidth consumption to exceed configured bounds and potentially making the hosting process unavailable. The fix validates peer-supplied label maps before they are retained or disseminated. Availability only; no confidentiality or integrity impact was identified.
github.com/temporalio/tchannel-go did not reject TChannel call fragments containing checksum metadata but no length-prefixed argument chunks. The fragment reader left its chunk slice empty and then unconditionally selected the first element. A network peer can supply such a malformed call fragment, including as a direct initial call request after completing the standard initialization handshake. On that inbound path, the resulting unrecovered Go slice-bounds panic occurs on a library-created dispatch goroutine and terminates the hosting process. This allows remote denial of service against applications that expose the listener to untrusted peers. The impact is limited to availability; no confidentiality or integrity impact was identified.
github.com/temporalio/tchannel-go did not validate the one-byte checksum-type field in inbound TChannel call frames. A network peer that can reach a listener can complete the standard initialization handshake and send a call request with an unsupported checksum type. The parser uses that value as an index into a four-entry checksum pool, causing an unrecovered Go panic on the connection read goroutine and terminating the hosting process. This allows remote denial of service against applications that expose the listener to untrusted peers. The impact is limited to availability; no confidentiality or integrity impact was identified.
temporalio/sqlparser accepts SQL containing deeply nested unary expressions and can return a correspondingly deep abstract syntax tree without enforcing an applicable nesting limit. The library's String and Walk operations recursively traverse that tree. An application that parses attacker-controlled SQL and later formats or walks the returned tree can encounter a runtime-fatal Go stack overflow that terminates the process; Go panic recovery cannot contain this condition. Temporal Server passes caller-controlled query input through the affected parser in archival, visibility, and worker-query paths. In affected validation paths, the Server recursively formats an invalid expression while constructing an error. In a supported authenticated deployment, a caller with namespace read permission can terminate the receiving Frontend or Matching process. The dynamically confirmed ListWorkers route additionally requires at least one retained worker heartbeat. Repeated requests can sustain a denial of service. The issue affects availability only; no confidentiality or integrity impact was identified.
temporalio/sqlparser can panic when Parse, ParseStrictDDL, or ParseNext processes a MySQL version comment whose contents are empty or consist only of one to five decimal digits. ExtractMysqlComment does not check the -1 result returned by strings.IndexFunc before using it as a slice boundary. The resulting Go runtime panic propagates unless the caller recovers it on the parsing goroutine, so applications that parse attacker-controlled SQL can terminate. Temporal Server exposes the affected parser through ListWorkers. When that API is enabled, an authenticated caller with namespace read permission can submit a malformed query that terminates the receiving Matching process. Repeated requests can sustain a denial of service. The issue affects availability only; no confidentiality or integrity impact was identified.
Temporal Server did not bound the work performed while searching for a Schedule's next action time. An authenticated caller with namespace write permission could create or update a Schedule that combines a fine-grained cadence with an exclusion calendar that rejects every candidate time, causing the server to evaluate excluded candidates without a per-search work budget. This can consume excessive CPU in Frontend and Schedule worker components. A persisted specification can also cause its backing Schedule Workflow to repeatedly fail and retry, allowing CPU consumption to continue without additional requests until the Schedule is deleted or its backing Workflow is terminated. Repeated or parallel exploitation can deny service. The issue affects availability only; it does not expose or modify Workflow data.
When OAuth authentication is enabled and browser-facing TLS terminates at a reverse proxy that forwards the callback to Temporal UI Server over HTTP, affected versions derive authentication-cookie Secure attributes from the proxy-to-server connection. Temporal UI Server can therefore issue access-token cookies, and refresh-token cookies when provided by the identity provider, without Secure even though the browser completed login over HTTPS.
A victim who visits attacker-controlled content while a credential remains live may expose that credential only if the attacker can also steer traffic for the UI hostname, prevent the browser's HTTPS connection from succeeding, serve the hostname over HTTP, and read a later same-site plaintext request. A malicious website alone cannot read the cookie, and passive observation of a successful TLS connection is insufficient. Effective HSTS, a blocking HTTPS-only warning, or TLS re-encryption between the proxy and Temporal UI Server prevents the demonstrated disclosure path.
A recovered credential may be replayed within the victim's assigned permissions. Refresh-token replay additionally depends on the identity provider's issuance, expiry, rotation, and reuse-detection behavior.
The frontend gRPC server's streaming interceptor chain did not include the authorization interceptor. When a ClaimMapper and Authorizer are configured, unary RPCs enforce authentication and authorization, but the streaming AdminService/StreamWorkflowReplicationMessages endpoint accepted requests without credentials. This endpoint is registered on the same port as WorkflowService and cannot be disabled independently. An attacker with network access to the frontend port could open the replication stream without authentication. Data exfiltration is possible, but only when a configured replication target is correctly configured and the attacker has knowledge of the cluster configuration, as the history service validates cluster IDs and peer membership before returning replication data.
The fix was applied per release line: it is present in 1.28.4, 1.29.6, 1.30.4, 1.31.2, and 1.32.0 and later releases on each line. Releases 1.31.0 and 1.31.1 do not contain the fix and are affected.
Temporal Cloud is not affected.
A writer role user in an attacker-controlled namespace could signal, delete, and reset workflows or activities in a victim namespace on the same cluster. Exploitation requires the attacker to know or guess specific victim workflow ID(s) and, for signal operations, signal names. This was due to a bug introduced in Temporal Server v1.29.0 which inadvertently allowed an attacker to control the namespace name value instead of using the server's own trusted name value within the batch activity code. The batch activity validated the namespace ID but did not cross-check the namespace name against the worker's bound namespace, allowing the per-namespace worker's privileged credentials to operate on an arbitrary namespace. Exploitation requires a server configuration where internal components have cross-namespace authorization, such as deployment of the internal-frontend service or equivalent TLS-based authorization for internal identities.
This vulnerability also impacted Temporal Cloud when the attacker and victim namespaces were on the same cell, with the same preconditions as self-hosted clusters.