Nezha vulnerable to cross-tenant terminal/file-manager session hijack via WebSocket stream UUID without ownership checkGHSA-q6xx-5vr8-p898

Summary

In nezha v1.14.13–v1.14.14 and v2.0.0–v2.0.9, the WebSocket endpoints GET /ws/terminal/:id and GET /ws/file/:id authenticate the caller only by the presence of a valid stream UUID, with no ownership check tying that UUID to the user who created the stream. Any authenticated dashboard user (including a RoleMember) who learns a live stream UUID can attach to the session and gain interactive shell access or full file-manager control on the target server — i.e. cross-tenant RCE.

This was silently fixed in commit 6661d6a (2026-05-18, shipped in v2.0.10). At submission time no public CVE/GHSA covers this fix, so operators of v1.14.x and pre-v2.0.10 v2.x deployments have no signal that they are running vulnerable code.

Details

Stream allocation — service/rpc/io_stream.go (v2.0.9):

func (s *NezhaHandler) CreateStream(streamId string) {
    s.ioStreamMutex.Lock()
    defer s.ioStreamMutex.Unlock()

    s.ioStreams[streamId] = &ioStreamContext{
        userIoConnectCh:  make(chan struct{}),
        agentIoConnectCh: make(chan struct{}),
    }
}

No creator is bound to the stream.

Stream attach — cmd/dashboard/controller/terminal.go (v2.0.9):

// @Router /ws/terminal/{id} [get]
func terminalStream(c *gin.Context) (any, error) {
    streamId := c.Param("id")
    if _, err := rpc.NezhaHandlerSingleton.GetStream(streamId); err != nil {
        return nil, err
    }
    defer rpc.NezhaHandlerSingleton.CloseStream(streamId)
    // ... WebSocket upgrade and bidirectional pipe ...
}

The only authorization check is GetStream(streamId) — "does this UUID exist in the in-memory map". getUid(c) is never compared against the user who called createTerminal. The same pattern is present in fmStream(c) in cmd/dashboard/controller/fm.go.

Where the UUID leaks:

createTerminal returns the UUID to the legitimate client, which then opens wss://<dashboard>/ws/terminal/<UUID>. As a URL path component the UUID is exposed via:

• Reverse-proxy access logs (nginx, Caddy, Cloudflare).
• Referer headers when the page embeds external resources or error reporters.
• Browser history / bookmark sync.
• Frontend telemetry (Sentry, Bugsnag) breadcrumbs that include the WebSocket URL.
• Any shared-tenant or multi-operator log viewer.

Any authenticated user with access to one of these side channels can attach to a live session.

PoC

1. Deploy nezha v2.0.9. Add at least one server. Configure two accounts: admin (RoleAdmin, owns the server) and member (RoleMember, no access to that server).
2. As admin, open the web terminal for the server. The browser opens wss://<dashboard>/ws/terminal/<UUID>. Capture this UUID from the network inspector, server access log, or Referer header.
3. From a separate session logged in as member, open wss://<dashboard>/ws/terminal/<UUID> (same UUID). The member's WebSocket attaches to the same ioStreamContext because terminalStream only checks GetStream(streamId) — no ownership check.
4. The member can now read the admin's shell output and inject keystrokes, achieving shell-level RCE on the target server, with no visible signal to the legitimate session owner.

Same flow works against /ws/file/:id (file-manager hijack: arbitrary read/write on the target server's filesystem).

Impact

• Severity: Critical. Interactive RCE on a server administered by another user, with no audit signal to the rightful session owner.
• Attack complexity: Low. The attacker needs an authenticated dashboard account (which any RoleMember is) and one captured UUID from a side channel.
• Confidentiality / Integrity / Availability: all High. /ws/file/:id exposes arbitrary read+write on the target filesystem; /ws/terminal/:id is a full shell.

This is the same impact tier as CVE-2026-46716 (cross-tenant cron RCE) and arguably worse, because the entry point is a passively-leaked URL rather than an authenticated POST — attackers do not need direct dashboard interaction once the UUID is leaked through logs or telemetry.

Fix reference

Already fixed in master by commit 6661d6a ("fix(rpc): bind io_stream sessions to creator to prevent terminal/fm hijack"):

• CreateStream now accepts a creatorUserID uint64 and stores it on the ioStreamContext.
• New IsStreamAuthorizedForUser(streamId, userID, isAdmin) helper.
• terminalStream and fmStream call this helper before the WebSocket upgrade and before the defer CloseStream(streamId), so a rejected attempt does not tear down a legitimate stream.

Shipped in v2.0.10 (2026-05-19). The v1.14 line has not received a backport.

Why this advisory

The fix landed silently. The other May 17–21 fixes received public GHSAs (GHSA-99gv-2m7h-3hh9, GHSA-rxf6-wjh4-jfj6, GHSA-hvv7-hfrh-7gxj, GHSA-w4g9-mxgg-j532, GHSA-6x26-5727-rrm9, GHSA-4g6j-g789-rghm) covering cron RCE, AlertRule trigger, telemetry leak, notification SSRF, DDNS SSRF, and agent forge-results respectively — but none cover the terminal / file-manager session hijack. This advisory closes that gap so operators of v1.14.x and v2.0.0–v2.0.9 know to upgrade.

Recommended action

• Publish this GHSA so v2.x operators below v2.0.10 see the alert in their dependency scanners.
• Either backport 6661d6a to a v1.14.15 release, or mark the v1.14 line end-of-life in SECURITY.md so operators understand the support boundary.