Email contact@jay739.dev, or open a private security advisory. Please do not open a public issue for a vulnerability.

I will acknowledge within a week. This is a solo-maintained project, so there is no formal SLA beyond that.

A local command line program. There is no server, no daemon, no listening socket, and no hosted component. It reads files you name, contacts hosts you configure, and writes files you name.

Network. It contacts only hosts named in the profile's egress allowlist. That is checked before every connection, including each redirect hop, so an allowed host cannot bounce it elsewhere. A connect_to override is a second destination and needs its own entry. The allowlist is empty by default, which blocks everything, so a profile with no sources reaches nothing.

A URL typed directly on the command line authorises that one host for that run, on the reasoning that typing it is the consent. Redirects from it remain constrained.

No telemetry. No analytics, no auto-update, no licence check, no crash reporting. The only outbound requests are the ones your profile configures. sitediff/source.py is the single module that opens a socket; everything else is local. Verify with a packet capture if you like.

TLS. Certificates are verified by default. verify_tls = false exists for endpoints that legitimately present an untrusted certificate, and it is opt-in per source.

Reports. Page content reaching an HTML report is HTML-escaped, so a page containing a script tag cannot execute when the report is opened. Credential response headers (set-cookie, authorization, x-api-key and similar) are withheld from reports even if a profile names them in compare.headers.

Redaction. [[redact]] rules run last, after every other transformation, on both page content and response metadata. If your pages contain tokens, session identifiers or personal data, add redaction rules. There are no default content redaction rules, because a wrong guess would silently hide real findings. Reports become CI artifacts and get passed around.

• Credentials in profiles. A profile may carry an Authorization header. Profiles live in your repository, so treat one containing a credential the way you would treat any other secret: keep it out of version control, or use a placeholder your CI fills in.
• Where reports go. They contain page content by design.
• Permission to fetch. The tool will contact whatever you configure. That you are allowed to is your call, not its.

It does not detect whether you are on a particular network and will not refuse to run based on your location. Such a check is bypassable in a one-line patch, redundant with targets already unreachable from outside, and would require inspecting your network interfaces, which is more invasive than the problem it claims to solve. Enforce network boundaries at the network.

The latest release only. This is a young project; there is no backport branch.