Many of the security checkers in this repository come with a small command line tool, and each README suggests using it as a CI gate: run node check.mjs https://staging.example.com in your pipeline and let a non-zero exit fail the build. Three READMEs said the tool exits 1 "when ... a request could not be completed at all". Three tools did not do that.
What the red team found
Today's red-team run planted 24 faults in subdomain-takeover-check. Its tests caught 6 of them (25%). All 8 faults in the CLI's argument parsing, text output and exit codes got through, because no test ever ran the CLI. All 8 in the provider table got through too: nothing pinned which providers count as claimable, and nothing tested an S3 region with a digit other than 1.
While writing those tests we found two real bugs:
- A failed DNS lookup looked like a clean result. The report's
okfield counted findings only. If the DNS-over-HTTPS lookup timed out, the CLI exited 0 and the page showed the green "No subdomain-takeover risk found". A zone audit with a resolver outage would pass. - The finding named the wrong host. Since an earlier fix, the tool checks every name in the CNAME chain, because
bucket.s3.amazonaws.comresolves on to an internal name likes3-w.us-east-1.amazonaws.com. The finding still showed the last name as its target, and its text said that name "matches Amazon S3's own hostname pattern". It doesn't. The name you need to reclaim or delete is the bucket's, and that is what the finding now shows.
The same bug in two siblings
grep found the line process.exit(report.ok ? 0 : 1) in 19 checker CLIs. For most of them ok already includes "every request completed". We ran all 21 URL-taking CLIs against http://127.0.0.1:1/, a loopback port that refuses connections at once. Two exited 0: dotenv-exposure-check and backup-file-exposure-check. In both, ok was just exposedCount === 0. Every request had failed, so nothing counted as exposed, so the build passed. Their pages also showed a green verdict. They now exit 1, and the verdict says how many paths could not be checked.
The guard
platform/ci/check_unreachable_exit.py finds every check.mjs whose usage line takes a URL, runs it against the dead port and fails on exit 0 or a hang. It covers 21 CLIs in under a second. On the old code it reports the two siblings above. A new checker is picked up by its usage line, so nobody has to add it to a list.
After the new tests, all 18 escaped faults are caught: 24 of 24.
None of this is a security audit of your site. The checkers report what one set of requests saw, and their READMEs say what they do not check.