GHSA-63MG-CXM4-98Q3
Multiple dashboard authorization, tenant isolation and resource exhaustion vulnerabilities
Ciclo de vida de la vulnerabilidad
| Fecha | Fase | Evento | Detalle |
|---|---|---|---|
| Divulgación | Publicada | GHSA-63MG-CXM4-98Q3 divulgada | |
| Monitoreo | Detectada | Ingresó al índice | |
| Monitoreo | Verificada | Última verificación contra las fuentes |
Summary
Multiple authorization, tenant isolation, authentication and resource-exhaustion defects were confirmed in the Dashboard. The fixes are available in Dashboard v2.6.0, which embeds admin-frontend v2.6.0. Deployments with untrusted members or member-operated agents are particularly affected; some resource-exhaustion paths are reachable without authentication when the corresponding endpoint is available.
The report was evaluated against Dashboard commits `15fa673c` and `686e04da`. Upgrade to Dashboard v2.6.0 or later. The affected-version range is conservative; individual features were introduced at different times.
Fixed releases
The fixes are released in Dashboard v2.6.0 https://github.com/nezhahq/nezha/releases/tag/v2.6.0. The companion administration UI is released in admin-frontend v2.6.0 https://github.com/nezhahq/admin-frontend/releases/tag/v2.6.0 and is included in the Dashboard release. The security changes include 027bfcf6 https://github.com/nezhahq/nezha/commit/027bfcf6521f192635b946b7583c3ab39500ab98 and the reviewed follow-up fixes in 7cd0371b https://github.com/nezhahq/nezha/commit/7cd0371b56f7bce83c3953b2bf4711b2105d5818.
Follow-up regression testing also fixed the real OAuth binding route's session parsing, reads of uncommitted public-host configuration, circular waits between task sends and transfer rollback, control leases released before a timed-out transport actually stopped sending, and missing final transfer admission for Agent upgrades and alert-triggered commands. The final native two-repository review found no remaining actionable defects within its reviewed scope.
Local validation passed: full Go tests and race/shuffle tests, vet, build, gosec (no findings under the repository configuration), the standard eight-Agent stress scenario, 86 frontend unit tests and 10 browser functional tests. CI also passed on the pushed commits: Dashboard Linux/Windows/macOS tests and builds, Linux race and security checks, eight-Agent stress, frontend CI, Chromium E2E, and CodeQL.
Impact and fixes
| Area | Confirmed defect and impact | Remediation |
| --- | --- | --- |
| NAT routing | A member could claim an undeclared Dashboard hostname and redirect requests, including login bodies, to their backend. Existing credential-header filtering does not protect a login body. | Member domains require an operator-configured exact allowlist and a declared Dashboard host. Reserved hosts and live ownership are checked on every route lookup, including disabled profiles. |
| DDNS | Unknown profile IDs passed binding checks. A later administrator-owned profile with that ID could be used by the member's server to update its DNS records. | Reject unknown IDs and require matching ownership at worker execution; administrator profiles are no longer implicitly shared. |
| Profile/password | An authenticated profile update with the correct current password and an omitted/empty new password replaced the hash with an empty-password hash. | Empty optional fields preserve their current values; nonempty passwords require 8–72 bytes. Login rejects empty credentials. |
| NAT request resources | NAT buffered the complete request body before forwarding, allowing unauthenticated memory exhaustion through an available NAT route. | Stream with backpressure, a 64 MiB body limit, bounded concurrency and connection lifetime. Safely detach the HTTP body framing when hijacking; reject HTTP/2 before dispatch. |
| Legacy OAuth login/binding | Binding state was not tied to its initiating user/session, could be replayed locally, and malformed or failed userinfo responses could yield an empty identity. | Require an initiating session, bind state to user and session, consume it atomically within five minutes, and validate status, JSON, identifier and a 1 MiB response limit. |
| Scheduled commands | A PAT holding cron write scope could create or modify executable schedules without cron execution scope. | Require `nezha:cron:exec` as well as the existing write scope and server restrictions. Provisional schedules cannot run before database persistence succeeds. |
| Server transfer | The destination became the owner before credential rotation was acknowledged, without a uniform gate on sensitive operations. Old NAT profiles also retained access after ownership changed. | Coordinate sensitive admission with ownership transitions; block non-admin control during transfer/recovery, revoke streams, check queued-task ownership at dispatch, and permanently invalidate previously issued file URLs on transfer. |
| Agent-controlled availability | Non-finite metrics/latency could break JSON and cause a WebSocket retry loop. One blocked task stream could hold up global dispatch. IP strings were used without family/format validation. | Normalize non-finite values, back off serialization failures, isolate dispatch with bounded queues/deadlines and reconnect stalled streams, and validate reported addresses. |
| Persisted configuration | Oversized service intervals could persist despite an API error and break startup. Alert history and old calendar-cycle traversal could consume excessive resources. | Enforce interval/history bounds, skip invalid historical service rows, quarantine unsafe alert rules, and bound calendar calculations. |
| Login/session handling | Form login admitted cross-site form submissions. Validly signed but revoked/stale sessions could trigger shared-IP bans. | Require JSON and same-origin browser requests; treat invalidated sessions as unauthenticated without that IP-ban path. |
| MCP and administrative integrity | Disabled MCP still wrote anonymous failure audits; missing source IP bypassed throttling. Downloads lacked a dedicated spool budget and a final HTTP write deadline. Logs could contain URL credentials, pagination could overflow, and user ID zero could be deleted by an administrator. | Early MCP kill switch, bounded audit worker and global/source quotas, 90-day audit retention, 32 download spools with write deadlines, credential-free request logs, safe pagination and reserved-user protection. |
Additional fixes preserve file-backed JWT signing keys across settings saves, deduplicate notification recipients, skip malformed theme metadata, preserve transfer history required by any applicable rule, clean up failed schedule registrations, and validate upload requests before consuming one-use URLs. The HTTP header timeout is also cleared when an HTTP/2 connection becomes active, preserving long-lived gRPC connections on Go 1.26.
Scope of verification and corrected interpretations
No complete numbered issue group was disproved. The following limitations matter when interpreting the original report:
- The empty-password change requires a valid session and the correct current password. It is not an unconditional unauthenticated password reset.
- Legacy OAuth account-binding abuse requires additional browser/cookie conditions. Reusable local state does not imply that an identity provider will redeem the same authorization code twice. This finding concerns legacy third-party login, not MCP OAuth.
- The transfer authorization gap was confirmed at task admission. The Agent already rejects configuration reads during its reload-timer window; end-to-end extraction of the previous owner's secret throughout the entire pending period was not demonstrated.
- Blocking propagation was confirmed with controlled task streams. We did not attack a deployment or measure real-world gRPC flow-control exhaustion thresholds.
- MCP transfers already had a 100 MiB per-file cap, context cancellation and IOStream quotas. The fixes add the missing spool and client-write bounds; the previous implementation was not wholly unlimited.
- Gin recovers the pagination panic; this was a request failure, not proof of process termination. Missing frontend index files were already skipped; malformed/missing metadata caused the startup problem.
Verification used source review and regression tests with local databases, simulated providers/streams and ordinary local HTTP fixtures. The standard eight-Agent compatibility suite validates normal operation and cleanup. No deployed instance was attacked and no exploit program is included.
Upgrade and operational notes
Read the upgrade guide https://github.com/nezhahq/nezha/blob/master/docs/security-upgrade.md before upgrading. In particular:
- Configure `dashboard_host` and `nat_member_domains` to retain member-created NAT routes. The updated administration UI exposes the allowlist as one hostname per line; omitted API fields preserve the existing policy and an explicit empty array disables member NAT.
- Replace implicit cross-user sharing of administrator DDNS credentials with explicitly owned profiles.
- Update automation PATs that create/edit schedules to include execution scope.
- Update form-based login clients and adjust invalid service/alert configurations. Reverse proxies must preserve the public request Host so same-origin browser login checks compare the correct authority.
The transfer gate does not erase old credentials or backups from an Agent host and cannot stop arbitrary OS commands that already started. Host handover still requires removing data that must not be shared. Trusted-proxy configuration, same-origin theme trust, the existing MCP file-read permission model and user-global Agent credential rotation are separate design/operational boundaries; they are not claimed as resolved by this advisory.
Thanks to @chaconneX for reporting these issues.
Incluye detalles técnicos publicados por el investigador. Úselos solo para validar y corregir sus propios sistemas.
Rangos afectados
Estado del parche por paquete
| Paquete / producto | Estado | Corregida en | Desde divulgación | Última versión |
|---|---|---|---|---|
gogithub.com/nezhahq/nezha |
Sin parche | — | — | v1.14.14 afectada |
Arsenal público
Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.
https://github.com/nezhahq/nezha/security/advisories/GHSA-63mg-cxm4-98q3Las referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.
