Summary

The V1 DurableTask stream handler registers worker-supplied durable task external IDs in an in-memory callback routing map before verifying that the authenticated tenant owns the task. If the tenant-scoped ownership check fails, the handler logs the error and continues, but the map entry persists until the stream closes.

Durable callback delivery resolves the destination stream by taskexternalid only, without checking tenant identity. As a result, a tenant A worker that knows a tenant B durable task external UUID can hold its stream open and receive tenant B's durable callback result payload when the callback is delivered on the same dispatcher process.

Impact

This advisory requires an attacker to successfully guess a durable task external UUID belonging to another tenant. Durable task external IDs are generated with uuid.New() (UUIDv4) and are not enumerable across tenants, so exploitation requires prior knowledge of a target task UUID through an out-of-band channel. Thus, while the following environments are impacted, there is an extremely low probability it would be exploited.

Who is impacted. Any Hatchet deployment that hosts more than one tenant on the same instance: - Hatchet Cloud (multi-tenant SaaS) - Self-hosted Hatchet with multiple internal teams / business units sharing one instance - Any deployment where a single tenant's API token can be obtained by an attacker

Single-tenant self-hosted deployments are unaffected in practice (the "victim" and "attacker" tenants would be the same).

Cross-tenant disclosure risk on DurableTask bidi RPC

This is a low-severity, low-risk cross-tenant data exposure vuln caused by blindly accepting a durable task id, in addition to a list of node and branch ids that identify records in that task's event log, and returning them to the caller via the handleWorkerStatus polling path.

Impact

This advisory requires an attacker to successfully guess a durable task's external uuid belonging to another tenant. External ids are generated uuid v4s (uuid.New() in Go), so exploitation requires prior knowledge of a target task UUID through an out-of-band channel. Thus, while the following environments are impacted, there is an extremely low probability it would be exploited.

Who is impacted. Any Hatchet deployment that hosts more than one tenant on the same instance:

Hatchet Cloud (multi-tenant SaaS) Self-hosted Hatchet with multiple internal teams / business units sharing one instance Any deployment where a single tenant's API token can be obtained by an attacker

Single-tenant self-hosted deployments are unaffected in practice (the "victim" and "attacker" tenants would be the same).

A flaw in ValidateOAuthState can allow unauthenticated OAuth state CSRF/login-CSRF due to an empty-state collision.

Summary

A missing authorization directive on the GET /api/v1/stable/dags/tasks endpoint caused Hatchet's tenant-membership check to be skipped for this route. A user authenticated to any tenant on the same Hatchet instance could query the endpoint with another tenant's UUID and a DAG UUID belonging to that tenant, and receive task metadata for that DAG.

This issue has been patched in v0.83.39. Hatchet Cloud has been patched and requires no action from users. Self-hosted users should upgrade.

Impact

Who is affected. Multi-tenant Hatchet instances reachable by an attacker who can obtain an account on that instance. On Hatchet Cloud, account creation is open by default. On self-hosted instances, the API must be reachable by the attacker and the hostname known; instances deployed inside a VPC or with signup restricted are not exposed to arbitrary external actors.

Prerequisites for exploitation. An attacker needed:

1. An account on the target Hatchet instance. 2. The victim tenant's UUID. 3. At least one DAG UUID (externalid) belonging to that tenant.

The two UUIDs are not treated as secrets — they appear in URLs, API responses, audit logs, invitation flows, shared run links, and dashboard screenshots — but an attacker does need to learn them through some out-of-band channel before exploitation is possible.

What could be disclosed. For each child task of a targeted DAG, the endpoint returned:

- displayname, actionid, stepid - workflowid, workflowversionid, workflowrunid, taskexternalid - tenantid, retrycount, status, timestamps - additionalmetadata (JSON)

The additionalmetadata field is the most sensitive: Hatchet workflows commonly use it to carry domain context such as user identifiers, customer IDs, feature flags, or correlation tokens. Its contents vary by deployment.

What was not disclosed. The raw task input payload is not part of this endpoint's response shape and was not exposed through this issue. The scope is limited to task metadata, not task arguments or results.

Exploitation status. We have no evidence that this vulnerability was exploited prior to the patch.

Root cause

Hatchet's multi-tenant authorization relies on an OpenAPI-driven middleware pipeline. Each authenticated operation declares x-resources: ["tenant", ...] in its spec. The populator middleware reads the declared resources, looks up the corresponding entities from request parameters, and stores them on the request context. The authz middleware then verifies that the authenticated user is a member of the tenant found on the context.

The listTasksByDAGIds operation accepted a tenant UUID as a query parameter, but its OpenAPI definition did not declare x-resources: ["tenant"]. As a result:

1. The populator, which early-returns when no resources are declared, did not populate the tenant onto the request context. 2. The authz middleware, which runs its membership check only when a tenant is present on the context, silently passed the request through. 3. The handler read the tenant UUID directly from the query parameter and used it as the filter in the downstream OLAP query.

The SQL query itself correctly filters by tenantid, so it returned only rows matching the supplied UUID — but the UUID came from the caller rather than from an authorization-validated context, so the filter bounded the response to the attacker-named tenant rather than to a tenant the caller was authorized to read.

Every other authenticated operation in the same path file (tasks.yaml) correctly declared x-resources. This endpoint was the only authenticated operation in the file that did not.

Patch

The fix adds the missing resource authz checks inline on the handler, enforcing valid tenant membership before the handler runs.

Shipped in v0.83.39.

Remediation

Hatchet Cloud. No action required. The patch was deployed on April 23, 2026 within the same day it was reported.

Self-hosted — recommended. Upgrade to v0.83.39 or later.

Self-hosted — if you cannot upgrade immediately. Either of the following reduces exposure until you can upgrade:

- Restrict account creation by setting SERVERAUTHRESTRICTEDEMAILDOMAINS to an allowlist of domains you control. This prevents arbitrary users from registering an account on your instance, which removes the most common path to the prerequisite account. - Ensure the Hatchet API is not exposed to untrusted networks. We generally recommend running Hatchet inside a VPC and fronting the API with authenticated network controls; deployments configured this way were not reachable by arbitrary external attackers.

Timeline

All times April 23, 2026.

- 14:05 — Reported to Hatchet. - 16:28 — Patch deployed to Hatchet Cloud and released as v0.83.39. - Public disclosure — this advisory.

Credit

Reported by @sajdakabir.

Hatchet thanks the reporter for responsibly disclosing this issue and for the clear, reproducible writeup.