v0.0.183
Released Nov 16, 2026
Release Notes
Four new probing security rules check what a site hands to anyone who asks: open JSON APIs, sensitive JSON fields, exposed files such as .env and .git, and auth endpoints that leak the user record. Four more catch mixed locale conventions, inconsistent terminology, copy whose own date has passed and broken byline conventions. Large audits peak at about half the memory, and a probe that never got an answer is now reported as not checked instead of clean. This release adds 8 rules: 336 rules on by default, 28 of them in Security, 29 in Content, 16 in E-E-A-T and 4 in Internationalization.
Added
- New rule
i18n/locale-mixing: one page that follows two locale conventions. It flags US and UK spellings of the same word (color/colour, center/centre, license/licence, organize/organise and more), day-first and month-first dates on one page (a failing check, since 3/4/2026 cannot be resolved), a bare$next to USD, CAD, AUD or NZD amounts, and spellings that contradict the page'slang(UK spellings underlang="en-US"). Pages whoselangis not English are skipped, so translations on a multilingual site are not judged. (squirrelscan/repo#1359) squirrel audit -f jsonnow carriescacheStats(squirrelscan/repo#1981) when the audit reused any page or sub-resource, so a fully cached re-audit is distinguishable from a fresh one. The field is absent on a cold audit. Added in JSON schema 1.21; this release ships 1.22.squirrel audit -f jsonchecks inissues,infoChecksandnotCheckednow carry optionalobservationmetadata (squirrelscan/repo#2302):mode(static, crawl, browser or field),applicability(applicable, not-applicable or unknown) and an optionaluncertaintythat qualifies the result. JSON schema version is 1.22.ax/agent-blockingcan probe every AI token (squirrelscan/repo#3788). Set[crawler] agent_access_all_tokens = trueto fetch the homepage as 17 AI bots that send their own User-Agent (GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, Perplexity-User, CCBot and more), about 15 extra homepage requests. Each probe records its status andX-Robots-Tag, and oneagent-access-probesinfo check lists them. It is off by default and the default request count and timing are unchanged. Verdicts now sayUA probe, because a spoofed User-Agent is not the real bot, and a blockedChatGPT-User,Perplexity-User,Meta-ExternalFetcherorMistralAI-Userfails likeClaude-Userdoes.- New rule
content/terminology-consistency(squirrelscan/repo#1368). Flags pages that call the same thing by a different name than the rest of the site: sign in / log in / login / signin, email / e-mail, sign up / signup / register / create account, FAQ / FAQs / Faq, and the casing of the product name (fromog:site_name). An agent asked to "log in" cannot reliably match a "Sign In" button when the site uses three labels. Spellings are counted per page and summed site-wide; a minority spelling is reported against the dominant one only when the dominant one has at least 70% of a concept's uses (at least 10 uses on 3 pages). The rule skips a crawl under 10 pages or with no clear norm, warns when one or two concepts are inconsistent and fails at three or more. It is a normal local rule: no model, no account, no surcharge. The per-page counts are stored in one bounded column (page_features.terminology_variants, schema version 43), so the streaming and legacy paths give the same answer. security/unauthenticated-api, a probing rule for API endpoints that hand out data with no login (squirrelscan/repo#1955). On anactiveoraggressiverun it sends oneGETwith no cookies, noAuthorizationheader and no query string to each same-origin API endpoint the audit found in your pages and scripts, at most ten, and never to a path that only takesPOST,PUT,PATCHorDELETEor that names a session or state-changing action. A 2xx JSON answer that carries records is graded by its field names: credential, personal or internal fields fail, user-identity or unclassified fields warn, and published content such as titles, slugs and prices is informational. The finding names the endpoint, the record count and the field names, never a value. A429, a bot or WAF challenge, a5xxor no endpoint to probe is reported as not checked, never as a pass; a401or the app's own403counts as protected. On apassiverun it sends nothing. See security/unauthenticated-api.- New probing rule
security/json-sensitive-fields(squirrelscan/repo#1956): on anactiveoraggressiverun it sends one cookie-lessGETto each same-origin API endpoint the discovery pass found (GET or method-less call sites only, at most ten) and parses the JSON. Password hashes, server-side keys the leaked-secrets patterns recognise, and tokens, secrets, SSNs, card numbers or CVVs under fields named for them fail; email addresses, phone numbers, addresses or dates of birth for five or more records warn. Values are never reported: items carry the field path, the kind of value and its length. - The entity map published to the cloud now carries a compact per-page
declaresindex (squirrelscan/repo#2117), so a hosted comparison can prove a gained@idreached every page that declared the old entity. When the index is clipped to its byte budget the map setstruncated.pagesTruncatedand comparisons report the change as unproven. security/exposed-fileschecks for well-known sensitive files a site should never serve:.gitand other version control metadata,.envfiles, configuration backups such aswp-config.php.bak,server-statusandphpinfo()pages,.DS_Store, backup archives, and database dumps (squirrelscan/repo#1932; Spring Boot Actuator paths are not probed). It is a loud probe, so it runs only with--probe aggressiveor--pentest; below that it sends nothing and the report lists it as not checked, with how to turn it on. It requests a fixed list of 20 paths on the audited origin only (aGETeach), never follows a redirect, reads at most 16 KB of each answer and stops at the first429or bot challenge, reporting the check as incomplete. A hit is confirmed by the file's content (aref:line, a[core]section,KEY=VALUElines, a magic number, a dump header), never by its status code, and a catch-all page is dropped against a random-path baseline. The finding names each exposed file and what kind of data it reveals, never a value from it. Two new coded skip reasons,probe-level-too-lowandprobe-incomplete, list probing rules that did not run in full under Not checked. See security/exposed-files.- New rule
eeat/byline-consistency(squirrelscan/repo#1364): reports pages that break the site's own byline convention. Pages are grouped by page type, and when a visible author byline or a visible date is present on at least 70% of a type's pages, the pages of that type without it are reported, with the convention they break (for example2 of 24 article page(s) have no visible author byline, which 22 (92%) of article pages carry). Only presence is a convention, so a site that never uses bylines is never flagged. It skips a crawl under 10 pages, page types with fewer than 8 pages, and sites where no type reaches 70%. A page it reports is no longer counted again byeeat/author-bylineoreeat/content-dates(they record how many they left out indetails.deferredToBylineConsistency), unlesseeat/byline-consistencyis disabled. See eeat/byline-consistency. - New rule
content/stale-dated-copy(squirrelscan/repo#1357): flags visible copy whose own date has passed, such asComing in 2024,New for 2025, a sale or event end date that is behind us,register nowbeside a past<time datetime>, and a countdown aimed at a past timestamp. Past-tense framing ("in 2019 we launched ...") reads as retrospective and does not fire, and noindex and archive pages are skipped. The date is judged by calendar month and can be set with thenowoption. The rule cache now also keys on the UTC month so a cached verdict is never replayed into a later month. - New probing rule
security/auth-endpoint-disclosure(squirrelscan/repo#1968): with--probe activeoraggressive, it sends one GET with no cookies or credentials to each same-origin current-user and session endpoint (the ones the audit found, plus/api/me,/api/session,/api/auth/sessionand/api/user, at most six) and fails, at error severity, when the JSON answer carries a password, a password hash, a credential secret such as a reset token or OTP secret, or the internal user record. Findings name the JSON path of each field, never its value. Login, sign-up, password, token and callback paths are never probed, and nothing is sent on a passive run. It shares one anonymous GET per endpoint withsecurity/unauthenticated-api, so an exposed/api/meis requested once and reported once, by this rule. A429, a bot or WAF challenge, a5xxor no endpoint to probe is reported as not checked (probe-incomplete,no-endpoints), never as a pass.
Changed
- Large audits peak at about half the memory they did. The CLI now forces one garbage collection on each side of building the report, so the report and its JSON output no longer stack on top of garbage left by the crawl and the rules. A 10,000-page benchmark audit peaks at 3.4 GiB of resident memory instead of 6.2 GiB, with the same wall time and the same report (squirrelscan/repo#1991).
SQUIRREL_STREAM_PHASE_MEM=1now also prints a line at the end of every audit phase, and on Linux each line shows the peak resident size since the previous one. NewSQUIRREL_HEAP_SNAPSHOT_DIRwrites a heap snapshot at those boundaries for diagnosing memory use. See Debugging.- When a report cannot be published because it is over the 1MB publish limit, the CLI no longer suggests auditing fewer pages (published reports are capped to fit, so that advice was wrong). It now says the report could not be capped to fit and that the full detail stays in your local project (
squirrel report). - Release binaries are built with
NODE_ENVpinned to"production"(bun build --production --define process.env.NODE_ENV='"production"'), so test-only knobs gated onNODE_ENV === "test"are compiled out whatever the build environment exports (squirrelscan/squirrelscan#1854, #1855). Bundled React now uses its production build for the HTML report; report output is unchanged.
Fixed
- Audit prompts only appear when both stdin and stdout are a terminal, so piped output never waits on a prompt (squirrelscan/repo#1171).
squirrel report --diffno longer calls a capped item fixed (#1781), and no longer counts findings on pages the new crawl did not reach as removed (#1835).squirrel analyzesends no network requests;security/http-to-httpsreports not checked there (squirrelscan/repo#1269).- The npm postinstall warns about memory when its self-install is killed by a signal (squirrelscan/repo#1675). The installer skips
sudowhen it is not present, and links the quickstart's musl note. - Fewer false positives:
local/nap-consistencyignores Event organiser and Person phones (#1877);a11y/label-content-name-mismatchjoins block children (#1876);a11y/identical-links-same-purposegroups by accessible name (#1875);a11y/link-in-text-blockskips icon-only and link-only blocks (#1874);a11y/color-contrastreads backgrounds from large utility stylesheets (squirrelscan/repo#4433);a11y/th-has-data-cellsends a row span at its row group;a11y/link-textresolves against the final URL (squirrelscan/repo#3366);content/reading-levelskips code blocks (squirrelscan/repo#1639). perf/bad-cachingparsesExpires, treats a past or zero date as no lifetime, and judges it against the responseDateor the fetch time, not the local clock (squirrelscan/repo#2243).- A scheme-less
Sitemap:line in robots.txt is read as a host only when it has a public-suffix TLD (squirrelscan/repo#2318), and a root path with a content query string is no longer classified as the homepage (squirrelscan/repo#3363). crawl/sitemap-lastmod-driftday thresholds must be non-negative integers (squirrelscan/repo#1680).- A probe that never got an answer is now reported as not checked, never as absent (squirrelscan/repo#1760). A 5xx (or any non-404/410 error status) on
/llms.txtor the Markdown probe no longer reads as "no llms.txt" or "no Markdown response".crawl/sitemap-coverageno longer lists pages as missing from a sitemap that was truncated or only partly read, andintegrity/orphan-pageno longer escalates a hidden page toward "likely injected" from such a sitemap (it is reported for review only). security/http-to-httpsnow runs: the CLI never gave the rule an HTTP client, so every probe failed silently and it always reported "no redirects" (squirrelscan/repo#1269). It now sends oneHEADrequest per sampledhttp://URL through the CLI request layer (user agent, timeout, per-hop egress checks) and fails the URLs that answer over plain HTTP without redirecting to HTTPS; a site that redirects passes. The probe time budget is the newprobeBudgetMsrule option (default 20000, next tosampleLimit), and a run that hits it sayspartial coverageinstead of reporting a clean result. With--offlinea redirect to another host is not followed. See security/http-to-https.- A page of deeply nested links no longer holds an audit in the parser (#1844). Parsing 20,000 directly nested anchors took 45 s, and now takes about 0.2 s. A link's text now stops at a link nested inside it, as a browser reads the page (the browser closes an open link when the next one starts): the inner link's text belongs to the inner link alone and is no longer repeated in every link around it. Pages without nested links parse as before.
--max-durationnow also bounds the start-URL redirect check; a silent host no longer runs the audit past its budget (#1853).
Security
- A start URL that redirects to a cloud metadata or link-local address no longer makes the CLI request it. The reachability check and the start URL redirect check, both run before the crawl, followed the start URL's redirects natively whenever the site's own hostnames were not checked (behind an HTTP proxy from the environment, or with
SQUIRREL_NO_HOST_PINNINGset), and the redirect check also followed a meta refresh or script redirect from the start page. Every hop now goes through the same check as the crawl, so a redirect to169.254.169.254,fe80::/10or a metadata hostname is never sent. The start URL is then reachable and audited itself, where before a refused redirect stopped the audit with "Cannot reach". Redirects within the site, andlocalhostor LAN start URLs, resolve as before, and a WAF challenge on the page an allowed redirect lands on is still detected (squirrelscan/repo#3336). See follow_redirects.
Known limits
- Known issue: on a re-audit where every page comes from cache,
security/unauthenticated-apiandsecurity/json-sensitive-fieldsfind no API endpoints to probe and report not checked (no-endpoints), never a pass. Run with--refreshto probe them again. Tracked in squirrelscan/squirrelscan#2081.
Update to this version
squirrelscan will auto-update, or run this command to update now: