How results are judged
- Only what was actually observed in this test is recorded; data we can't get is shown as "No data" or "Could not read" and is never filled in from other sources.
- Checks that weren't run, failed or lacked information are never shown as "OK" or "no leak".
- Results reflect the network behavior observed in this test and don't guarantee the same at other times or for every AI service.
Exit IP
The page makes one request to this site; the server reads the public address used for that request and looks up its country and region, city, ASN, network owner and network type.
What it can tell you
The exit your browser uses to reach this site.
What it can't tell you
With split routing, other websites may be reached through different exits. Location and network type are estimated from IP registration data and may be inaccurate.
AI service connection check
The browser requests the echo address on the ChatGPT, Claude, Grok and Copilot domains to read the exit IP each service sees and the edge location it connects to, and records the time taken. Gemini has no readable echo, so we send one opaque request to prove it's reachable and record the time; its exit IP, location and edge location show as "not provided". This site also resolves each service's domain to get the target IP.
Location: the country or region of the exit seen by that service; where an official list exists, we also check whether it's supported. OpenAI and Anthropic publish full lists; Microsoft only says it's available worldwide except mainland China and embargoed markets; xAI (Grok) hasn't published a list, so no official-support judgment is made — we only show "Reachable in our test" or "Not reachable in our test".
The home page cards show the exit IP each service sees; when it differs from this site's, it's highlighted with its location, and a summary above says whether split routing exists.
The lines on the home page globe start from your exit's real location; the endpoints are only an illustrative animation and don't represent where each service's servers actually are.
What it can't tell you
Whether you can sign in or use every feature, and whether APIs or command-line tools (such as Claude Code) use the same exit.
WSS long-lived connection
The browser opens a WebSocket connection to this site, and the server assigns a separate session to each test.
- Continuous: regular heartbeats and echoes (returned after the server actually processes the message), plus simulated 20-chunk AI streaming output. Shorter tests run at a faster pace: at 1 minute, a heartbeat every 5 seconds and an echo and streamed output every 15 seconds; at 5 minutes, a heartbeat every 10 seconds, an echo every 30 seconds and streamed output once a minute; at 30 minutes, a heartbeat every 10 seconds, an echo every 30 seconds and streamed output every 2 minutes.
- Resume after idle: after connecting, the line stays silent and messages are sent at minutes 1, 2, 4, 7, 10, 15, 20 and 25 to check whether the connection can still communicate.
A reply taking more than 10 seconds counts as a timeout; a closed connection counts as a drop and is reconnected automatically (up to 5 times). These intervals are current settings and will be tuned based on real-world measurements.
Please note
Keep the page open during the test. Periods when the page was in the background are marked in the log, and results from them may be unreliable.
HTTPS / SSE streaming
The browser requests this site's streaming endpoint, and the server sends numbered chunks continuously for the selected tier (Quick 40 seconds and Standard 2 minutes, one chunk every 250 ms; Deep 5 minutes, one chunk every 500 ms), followed by an end marker; optionally it goes silent for 30 seconds midway to simulate AI thinking. The page records time to first data, arrival gaps, the longest stall (a gap over 1.5 seconds counts as a stall; 3 seconds for Deep), whether sequence numbers are continuous, and whether the end marker arrived.
What it can't tell you
The test uses data sent by our test endpoint and doesn't reflect how fast AI models generate text.
DNS leak
For each test, this site generates a set of random domains (2 for a quick test, 10 for a deep test). When the browser visits them, the lookups follow the DNS path you actually use and eventually reach our own authoritative DNS server, which records the resolver addresses that sent them. Whether the web requests themselves succeed is not used as evidence.
Results are grouped by operator and region and compared with the country or region of your web exit. Public DNS such as Google and Cloudflare runs on anycast networks, so the registered location doesn't reflect where lookups are actually handled; we don't judge leaks on that basis.
What it can't tell you
Currently only the IPv4 resolution path is covered. A DNS exit that differs from your public IP doesn't necessarily mean a leak — judge it together with your network and proxy setup.
WebRTC
Through WebRTC, the browser collects candidates from Cloudflare's and Google's public STUN services. This site first resolves the STUN server names to addresses so that proxy apps rewriting DNS don't interfere. Only actual, non-relay public addresses count as evidence and are compared with your web exit; local addresses hidden by the browser (mDNS) aren't recorded.
What it can't tell you
The test can only show what the browser allows it to read and can't guarantee finding every address exposure.