v0.0.136

The CLI now checks the audited site's own hostnames before it connects, so a name that resolves to a cloud metadata or link-local address is never requested. Set SQUIRREL_NO_HOST_PINNING to turn the check off on a network where connecting by address breaks routing.

Security

• The CLI checks the site's own hostnames before it connects. The audited host, its www twin and the domains in [project] domains went out by name with no address check, so a site that controls its DNS, or rebinds it mid-audit, could point one of them at a cloud metadata service and the audit would request it and store the answer. The CLI now checks every answer for them and connects to the address it checked, for pages, redirect hops, root files, the start URL's reachability check and redirects, assets and the end-of-crawl probes, in squirrel analyze too; a link-local or metadata answer is never requested (a page is recorded as failed with "Not requested", an asset as refused). Two rule probes still go out by name: the security/http-to-https HEAD request and the GraphQL probes of an active or aggressive probing intensity. Behind an HTTP proxy from the environment, or with SQUIRREL_NO_HOST_PINNING set, every request goes out by name as before. See follow_redirects.

Docs

• compare_audits says when a rule could not be checked. A rule the newer audit could not run is reported as not_checked, not as resolved. See MCP.
• The rule-count tables match the 311 rules. The README and the docs index list Images 16 and Content 25.
• The AI-agent guide no longer states a tool count that goes stale.