Cloud scans
Run a scan against something you do not host on your machine (a vendor’s site, a host’s AI services, a remote MCP server) and know exactly what Keryn’s servers ask of it.
Before you start
Section titled “Before you start”- A Keryn Cloud account. Open Cloud in Keryn Web and select Sign in. Sign in on the account page that opens and approve the code shown in Keryn Web. The Cloud tab continues automatically after you approve it.
- The address of the target, as a public host name. IP addresses are refused.
- Read what leaves your machine: a scan sends the address you type.
- In Cloud, choose Trust site, AI surface or MCP server.
- Type the address (
https://vendor.example,ai.example.com,https://mcp.example.com/mcp), or pick one under Try an example, and press Scan. - The result opens in the results view. The history below the form lists earlier scans, which you can reopen or delete.
A scan finishes inside 45 seconds. A target that answers too slowly ends the scan with “the target answered too slowly”.
Trust site
Section titled “Trust site”Keryn’s servers read the site as an anonymous visitor would and fetch the most useful pages first: trust centre, security, privacy, sub-processors, DPA, SOC 2 and ISO reports, AI policy and model cards before anything else. To find them, the crawl reads robots.txt and up to five sitemaps, /.well-known/security.txt, and the usual trust paths, and it tries the organisation’s own trust., security., trustcenter. and www. hosts. It follows a link to a hosted trust-centre platform (such as Vanta, Drata, SafeBase or Whistic) only when the site itself links to it, and to at most two of them. Redirects within the organisation’s own domain are followed (for example from www.example.com/docs/security to docs.example.com). If the address you typed redirects once to the organisation’s own new domain (for example an old brand domain), the scan follows that one redirect; no other redirect to another domain is followed.
The crawl honours robots.txt: the KerynCloud group if there is one, otherwise *. The address you typed and the sitemaps robots.txt names are always read; every other address it disallows is skipped and listed in the result.
Every scan is bounded: at most 48 pages and 160 requests, 40 pages from the site and 8 from any other host, three links deep, all inside the 45 seconds. Text is scored up to 30 KB per page and 400 KB per scan, trust pages first, and every page keeps at least 2 KB. The pages are scored by the Trust Scanner (the same engine behind the keryn trust scan scorecard). Trust scans use the default web ports (80, 443) of the host name; another port needs a verified domain.
How to read the result
Section titled “How to read the result”- The headline is “N of 48 trust artifacts evidenced”, a score from 0 to 100 and its band (25, 45, 65 and 85 are the band edges). Each artifact that counts cites the address it was found at.
- Present means the artifact’s own page, a heading, a page introducing itself as it, or a link whose text names it says so. A practice stated on its umbrella page (encryption on the security overview, retention in the privacy policy) counts too. When the evidence is a link, the result says the linked page was not read.
- Mentioned means a page only talks about it: shown with the page and the words, never counted.
- Not read comes first among the limitations: trust centres that the site links to but that build their content in the browser (several hosted platforms do) are listed as found but unreadable, not scored as empty.
- Blocked sites (HTTP 403, a bot wall or a sign-in wall on the address you typed) say so, show no score and report nothing as missing.
AI surface
Section titled “AI surface”The passive check requests, over HTTPS (HTTP if there is no HTTPS answer), the host’s front page and a short list of well-known paths: /v1/models, /api/tags, /api/version, /.well-known/mcp, /mcp, /sse, /openapi.json, /docs and /health. Responses are capped at 64 KB and redirects are not followed. What comes back is classified into AI services (model runtimes, UIs, gateways, workflow builders, notebooks, vector stores, MCP servers) with the same classifier as keryn surface classify, which is how exposure and missing authentication become findings.
The active check is an extra switch, Active port checks, that is off and disabled until the host is under a domain you verified. It then opens a plain TCP connection to eighteen ports where self-hosted AI services commonly listen (3000, 3001, 4000, 5000, 5678, 6006, 6333, 7860, 8000, 8001, 8080, 8443, 8501, 8888, 9000, 9090, 11434 and 19530) and, for each that is open, makes one HTTP or HTTPS request to /. Nothing is sent that tries to log in, guess a value or change state. A scan with active port checks costs 1 credit; the credit is held when the scan starts, the scan is refused if your organization does not have it, and a scan that fails is not charged. The site, host and MCP scans without the switch cost no credits.
MCP server
Section titled “MCP server”Keryn’s servers send the MCP handshake (initialize) to the URL, then read the server’s tools, prompts and resources lists (capped at 500 items each). They never call a tool. The tools are judged by the same engines as keryn mcp inventory, including the poisoning screen over names and descriptions, and the report adds the server’s transport and authentication posture (for example plain http, or no authentication challenge). Only servers reachable over HTTP are scanned; a server that starts as a local process cannot be scanned in the cloud, so use keryn mcp audit for those.
Website checks from Test app
Section titled “Website checks from Test app”For a public site you own or have permission to test, open Test app in Keryn Web. If the browser cannot fetch a target because the site does not allow cross-origin requests, the result can offer Run on Keryn Cloud. Sign in to Keryn Cloud, review the estimate, and confirm that you have permission to test the site before starting the run.
Cloud website checks support Response checks and Find more pages. They send ordinary GET requests to public pages and run without target credentials. Cloud stores the report. The run is limited to 25 URLs across three hosts, 100 requests, and 40 seconds. Link discovery stays within the starting hostnames and follows links one level deep. Test app blocks the Cloud fallback when the browser run or draft has custom path restrictions.
Examples to try
Section titled “Examples to try”Try an example under the address field fills in a target, labelled by who runs it:
| Kind | Example | Run by | What it shows |
|---|---|---|---|
| Trust site | Keryn demo trust centre | Keryn | A fictional company with security, privacy, sub-processor, DPA, SOC 2, AI policy and model card pages and a security.txt. |
| Trust site | Vanta (https://www.vanta.com) |
third party | A real public trust centre, read as any visitor would. |
| MCP server | Keryn demo MCP (https://api.keryn.app/demo/mcp) |
Keryn | Deliberately flawed: one tool with a poisoned description, one over-broad tool, one clean tool. It never runs anything: a tool call is refused. |
| MCP server | DeepWiki (https://mcp.deepwiki.com/mcp), Context7 (https://mcp.context7.com/mcp) |
third party | Public documentation servers that answer without sign-in. Only their lists are read. |
| AI surface | keryn.app |
Keryn | Passive checks only. |
The demo MCP server can also be judged offline, with no network at all:
keryn mcp inventory --exampleWhat you just proved
Section titled “What you just proved”The target saw ordinary requests from Keryn’s servers, nothing from your machine, and nothing was sent to Keryn except an address you typed. A scan that was refused (a private address, a port without a verified domain, an IP address) says why in the Cloud page.
- Verify a domain to allow the active checks.
- Limits, retention and deletion.