# Protocol: browser path leakage on Cursor-hosted USA cloud VMs

Frozen before measured collection, 17 November 2026, UTC.
Amendment of the 22 October 2026 Kenya / Proton / NetworkManager protocol.

## Decision and scope

Earlier FreePolicyTools studies measured VPN-website warning pages, tracker
payloads inside a working Proton CLI tunnel, HTTP-client environment-variable
routing, and paid-proxy CONNECT behaviour. The October 2026 browser-path draft
asked which Playwright Chromium paths still expose a non-intended public
address on a Kenyan research PC with Proton Desktop, NetworkManager WireGuard,
and Proxy-Seller endpoints.

This cloud amendment asks a narrower question suited to Cursor-hosted USA
datacenter VMs:

**When a fresh Playwright Chromium context is told to use a residential proxy,
do page-fetch IP echoes and WebRTC/STUN candidates follow that proxy path, or
does any observed surface still match the cloud-direct egress?**

Primary surfaces remain researcher-operated: a local leak page that gathers
WebRTC ICE candidates and fetches independent public IP echoes. Vendor “are you
protected?” pages are not the primary sample.

Out of scope for this amendment (explicitly removed):

- Kenya ISP wireless baseline
- Proton Linux CLI / Desktop / kill-switch conditions
- NetworkManager WireGuard profiles (`fpt-de` and siblings)
- Stacked VPN-plus-proxy conditions
- Any claimed CometVPN tunnel measurement without a separately proven working
  VPN path on this VM

CometVPN is VPN-only. It must not be used as HTTP/SOCKS. Affiliate and product
context may appear in the article; tunnel results require a proven path first.

## Runtime platform

- Cursor Cloud Agent VM, USA datacenter egress
- Playwright Chromium (isolated context; no personal profile)
- Locale `en-US`, timezone `UTC`, viewport 1360×900, JavaScript on, ordinary TLS
- Local fixture on `127.0.0.1` (loopback bypass for browser proxy)
- Public echoes: `api.ipify.org`, `api64.ipify.org`, `ipinfo.io/json`,
  `cloudflare.com/cdn-cgi/trace`
- WebRTC: `RTCPeerConnection` with `iceServers`
  `stun:stun.cloudflare.com:3478` only. Host, srflx and relay candidates
  recorded. No TURN.
- Duration: ≤7 days from freeze (≤14 only if an amendment justifies data volume)

## Frozen route conditions

### Active now

1. `cloud-direct` — no browser proxy; Cursor VM default egress. Collected first.

### ProxySeller residential (unlocked in A1)

Unlocked after Runtime Secrets appeared in process env. Do not invent credentials. Do not read secrets from git, chat, or committed `.env` files.

2. `proxy-us-http` — ProxySeller residential USA IPv4 as Chromium **HTTP** proxy.
3. `proxy-us-socks` — same residential package as Chromium **SOCKS5** proxy.

### Supplementary (after proxy legs unlock)

4. `proxy-dead` — invalid local proxy port (`http://127.0.0.1:9`); local page
   loaded with loopback bypass; public fetches and WebRTC observed
   (fail-closed / fail-open behaviour).

Pass 2 reverses main-condition order. A1 collected HTTP/SOCKS both passes plus `proxy-dead` once.

## Amendment hook A1 — unlock ProxySeller legs

**Trigger:** any process environment variable whose name matches
`/(?i)proxy|seller/` becomes present (values remain redacted). Prefer names
already documented by the Cloud Agents Runtime Secrets UI; do not invent new
credential sources.

**Discovery method used at freeze (negative result):**

1. `env | rg -i 'proxy|seller'` with values redacted via `sed -E 's/=.*/=<REDACTED>/'`
2. `printenv | cut -d= -f1` scanned for proxy/seller/socks/resid/comet/vpn tokens
3. Python `os.environ` name scan for the same tokens plus an explicit candidate
   list (`PROXYSELLER_HOST`, `PROXYSELLER_USER`, `PROXYSELLER_USERNAME`,
   `PROXYSELLER_PASSWORD`, `PROXYSELLER_HTTP_URL`, `PROXYSELLER_SOCKS_URL`,
   `PROXYSELLER_HTTP_PORT`, `PROXYSELLER_SOCKS_PORT`, and related aliases)
4. `cursor-cloud` `list-available-secrets` (publish-app secrets catalog — empty;
   distinct from Cloud Agents Runtime Secrets)
5. Dashboard environment page navigation reached Cursor authenticator login only
   (no authenticated secret-name read)

**Result at freeze:** 42 environment keys total; **0** matching
proxy/seller/socks Runtime Secrets. Named `PROXYSELLER_*` candidates: none
present. Therefore `proxy-us-http`, `proxy-us-socks`, and `proxy-dead` remain
**blocked / pending**.

**On unlock:**

1. Re-run the discovery commands above; record **names only** (never values) in
   `evidence/env-discovery.json`.
2. Rebuild `private/proxies.json` via `prepare-proxies.py` from discovered env
   vars (private; gitignored).
3. Collect `proxy-us-http` and `proxy-us-socks` for passes 1 and 2 (reversed
   order on pass 2), plus `proxy-dead` once.
4. Record an amendment note with timestamp in this file’s Amendment log.

Collectors must refuse to start a proxy condition if the required env vars are
absent. Missing secrets are a guard failure, not a “direct path” finding.

## Collection rules

- Isolated Chromium via Playwright; one context per visit.
- Route verification before and after each main condition via `route-check.py`
  (Python urllib echoes — not the research browser). Omit Proton/nmcli assumptions
  that do not apply on this VM; keep IP-echo route checks and basic `ip route`.
- Do not complete CAPTCHAs, log into third-party sites, or submit forms.
- Failed guards exclude that condition until reviewed; they are not successes.
- Never commit secrets, raw public IPs, or private screenshots.

## Public evidence

Raw addresses, credentials and original screenshots stay in `private/`
(gitignored). Public records keep only:

- whether a WebRTC candidate matched the browser document IPv4/IPv6 echo
- whether it matched the cloud-direct baseline or the proxy exit from the same
  condition’s route check (opaque aliases / private hashes)
- RFC1918/link-local/unique-local classification
- fetch success versus network error
- country/org labels from ipinfo when present, not the IP

## Interpretation limits

A candidate or document echo matching the cloud-direct baseline during a proxy
condition is evidence of a split or non-engaged browser path on this Chromium
context and this VM. It is not proof of provider logging, deception, a
product-wide defect, or named-person identification.

USA cloud datacenter egress is not a matched residential control for residential
proxy exits. Proxy findings are endpoint- and account-configuration specific.

CometVPN tunnel behaviour is **not** measured in this amendment unless a later
amendment proves a working VPN path on this VM (TUN device presence alone is
insufficient).

## Funding / disclosure

FreePolicyTools is a ProxySeller affiliate (`/proxyseller/`) and a CometVPN
affiliate (`/cometvpn/`). ProxySeller residential capacity used for this study
was purchased for FreePolicyTools cloud research. Affiliate links in the article
include disclosure. Measurements are researcher-operated fixtures / IP echoes /
document fetches, not competitive benchmarking of ProxySeller uptime.

## Amendment log

| ID | At (UTC) | Change |
|---|---|---|
| A0 | 2026-11-17 | Cloud freeze: replace Kenya/Proton/NetworkManager matrix with `cloud-direct` + pending ProxySeller HTTP/SOCKS. CometVPN context-only. |
| A1 | 2026-11-17 | Unlocked ProxySeller legs via Runtime Secrets names PROXYSELLER_HOST, PROXYSELLER_HTTP_PORT, PROXYSELLER_SOCKS_PORT, PROXYSELLER_USER, PROXYSELLER_PASS. Collected HTTP/SOCKS/dead with Playwright; SOCKS via local anonymizer because Chromium rejects SOCKS5 auth. |
| A2 | 2026-11-18 | Article gap deepen (same branch): publish orphan collector artifacts, document ICE timeouts / baseline document≠srflx / sha-relations vs cross-compare, add REPRODUCE.md and deepen site page. No new matrix visits. |
