/

CVE-2026-76073

Severity:

HIGH RISK

(8.8)

IDOR (CWE-639) in Label Studio Annotation API Endpoints

CVE-2026-76073

Deferred

Vulnerability type

Category:

Broken Access Control

CWE:

CWE-639

Impact:

Cross-Organization Data Access and Unauthorized Modification

Severity:

HIGH RISK

(8.8)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

Label Studio exposes annotation records through an endpoint whose database query is not restricted to the requesting user's organization. Any authenticated user who can identify other annotation IDs may be able to read, modify, or delete annotations belonging to another organization on the same instance.

Affected context

The vulnerability occurs in Label Studio annotation API paths that retrieve annotations by primary key without applying the active organization's scope. It is most relevant in multi-tenant deployments where multiple organizations share an instance and annotation identifiers are predictable or enumerable. The related annotation conversion path may expose the same unscoped-query behavior.

Exploit preconditions

  • Attack surface: Authenticated annotation and annotation-conversion API endpoints
  • Required access: A valid account on the affected Label Studio instance
  • Trigger point: Requests that retrieve, update, delete, or convert annotations by identifier
  • Why exploitation is possible: The endpoint uses an unrestricted annotation queryset, while permissions confirm authentication and named privileges without enforcing organization-level object ownership
  • Additional condition: The attacker must obtain or infer identifiers for annotations outside their organization

Why this happens

The underlying issue is an object-level authorization failure caused by separating permission checks from data scoping. The endpoint checks whether the requester is authenticated and has the relevant named permission, but its queryset does not constrain records to the requester's active organization. As a result, a primary-key lookup can resolve objects belonging to another tenant. This maps to CWE-639: Authorization Bypass Through User-Controlled Key, because an identifier supplied by the requester can select a resource without validating that the resource is within the requester's authorized scope.

Potential impact

  • Unauthorized users may read annotations from other organizations
  • Attackers could modify annotation content or workflow state
  • Attackers could delete annotations belonging to other tenants
  • Sensitive labeling data may be exposed across organizational boundaries
  • Annotation integrity and downstream training data could be affected

General mitigation guidance

Mitigation generally focuses on enforcing tenant and object ownership at the data-access layer rather than relying only on endpoint permissions.

Common defensive patterns include:

  • Scope every annotation queryset to the requester's active organization or equivalent tenant boundary
  • Apply the same organization constraint consistently to read, update, delete, and conversion operations
  • Perform object-level authorization after lookup and before returning or mutating a record
  • Treat client-supplied identifiers as untrusted selectors, not proof of ownership
  • Use least-privilege permissions and avoid granting broad annotation privileges by default
  • Prefer non-sequential or opaque identifiers as a defense-in-depth measure, while retaining authorization checks
  • Add isolation tests covering users, organizations, and all supported annotation actions

How teams detect issues like this

Detection can combine source review, automated analysis, and runtime policy checks:

  • Search for model querysets using unrestricted patterns such as Model.objects.all() in tenant-aware API views
  • Review endpoint lookup methods for missing organization, project, or ownership constraints
  • Flag authorization logic that checks authentication or named permissions without object-level validation
  • Test cross-organization read, update, delete, and conversion requests using identifiers from another tenant
  • Enforce code-review policies requiring tenant scoping for every multi-tenant data access path
  • Monitor audit logs for repeated sequential identifier access or cross-organization authorization failures
  • Neutral automated review and static-analysis tools may help identify unscoped querysets and missing authorization checks

Sources

SourceLink
github.comgithub.com/HumanSignal/label-studio
github.comgithub.com/HumanSignal/label-studio/blob/1.23.0/label_studio/tasks/api.py
github.comgithub.com/HumanSignal/label-studio/issues/9796
vulncheck.comvulncheck.com/advisories/label-studio-through-cross-organization-annotation-access-via-unscoped-annotationapi-queryset

Frequently asked

questions

Yes, where the affected API endpoints are reachable by authenticated users. Exploitation does not require network-level access to the server, but it does require an account on the instance.

Not necessarily. The described permission model allows authenticated accounts to satisfy the endpoint's permission checks, making organization-level authorization the important missing control.

Enforce organization or tenant scoping in the queryset and repeat object-level authorization checks before every annotation read or mutation.

Treat all resource identifiers as untrusted, require tenant-aware data-access patterns, test cross-tenant isolation in CI, and review every endpoint for authorization on individual objects rather than only on users or roles.

Find out what's already
exploitable in your codebase.