Skip to main content
Releases

v0.0.184

Release notes

Six new rules, including four probing security checks that knock on the doors a vibe-coded app forgets to lock: open redirects, Supabase tables readable with the public anon key, exposed Spring Boot Actuator endpoints, and admin APIs that trust the frontend to hide them. Batch audits from a URL list, squirrel support in the terminal, accessibility rules that name the WCAG criteria they test, and findings ranked by how common they are across sites.

Added

  • squirrel support (squirrelscan/repo#1090): open and follow threaded support tickets from the terminal. new, list, view <ref> and reply <ref>, each with --json; new takes --type, --subject, --message, --run-id, --rule-id, --report-id and --attach-doctor (the squirrel self doctor output rides in metadata, never the message, and is asked about first at a terminal). It needs squirrel auth login. When support has replied, audit, crawl and analyze print Support replied on <REF>: squirrel support view <REF> on stderr, read from a value cached in settings and refreshed in the background at most hourly. The local MCP server gains create_support_ticket, list_support_tickets, get_support_ticket and reply_to_support_ticket. See squirrel support.
  • squirrel audit --urls-file <file> --out-dir <dir> audits a list of URLs in one batch (squirrelscan/repo#1231): one report per URL in --out-dir, named by a hash of the URL, plus results.jsonl with one line per URL (status, reason, score, pages, exit code, timings). Re-running the same command resumes and skips URLs already done; --retry-failed runs the failed ones again and --restart starts over. --concurrency sets how many audits run at once, --url-timeout is a hard time limit per URL in seconds, --max-urls caps the list (100000 by default), and --allow-private-hosts allows loopback and private hosts, which are refused by default. The batch exits 6 when at least one URL failed, was blocked, unreachable, redirected, timed out, crashed, was rejected or was skipped. Batch mode is local only for now: signed-in batches are refused unless you pass --offline. See Batch audits.
  • New rule content/font-consistency (squirrelscan/repo#1376): flags font sprawl (more distinct primary font families site-wide than max_families, default 4), a family used on one page only, and @font-face families no font-family names. It reads <style> blocks and style= attributes only, because linked stylesheet bodies are not fetched yet, and every finding says so; a site with no inline font data is skipped, not passed. Icon fonts, generic keywords and system stacks are not counted, and a crawl under 10 pages is skipped. Local rule, no account, no credits. The per-page signal joins the site-dom-signals record, so each site's first audit after the upgrade runs fresh once.
  • squirrel audit --content-store <path> uses that file as the page-body store for the run, the same as SQUIRREL_CONTENT_STORE_PATH. The default store is shared by every project, so point it at a new path for a cold run. When bodies were already in the store, the console output says how many (the report formats do not carry the counts). See squirrel audit. (squirrelscan/repo#1914)
  • New rule security/open-redirect. A probing rule (runs at --probe active or above) that finds same-origin URLs whose query carries a redirect-shaped parameter (next, url, redirect, returnTo, continue and the like) in crawled page URLs, internal links and discovered GET endpoints, and requests each with that parameter set to an off-site value on a host that cannot resolve (squirrelscan-redirect-probe.invalid), first as an absolute URL and then protocol-relative. It warns when the 3xx Location points to that host, and never follows the redirect or reads the body. State-changing paths and non-GET call sites are never requested, robots.txt is respected below aggressive, and at most ten parameters are probed inside the shared probe budget. Findings name the page and the parameter, not the query string. That makes 337 rules, 29 of them in Security. (squirrelscan/repo#1936)
  • security/supabase-rls, a loud probing rule for Supabase tables anyone can read with the public anon key. It finds the Supabase project and anon key a site ships, then sends one HEAD /rest/v1/<table> with Prefer: count=exact to each table the site's own code queries, and judges only the status, the content type and the row total in Content-Range: a total above zero fails. It never sends a GET, never reads a response body and never guesses table names, so no row, value or column name is fetched or stored. It runs only with --probe aggressive, not with --offline, at most ten requests per audit, and never with a service-role or secret key. A 429, a bot challenge, an HTML 401, 403 or 404 page, a network error or a table left out by the caps makes the run not checked (probe-incomplete), never a pass, and a table meant to be public can be waived with [[ignore]]. See security/supabase-rls. That makes 341 rules, 32 of them in Security (squirrelscan/repo#1959).
  • New rule security/exposed-actuator finds Spring Boot Actuator endpoints that answer anyone (squirrelscan/repo#1961). When probing is active it sends read-only GET requests to /actuator, /actuator/env, /actuator/configprops and /actuator/mappings under /actuator and any same-origin actuator base the crawl found, and on an aggressive run one GET /actuator/heapdump that reads at most 64 bytes. An exposed server writes a heap dump when asked, so a run asks for at most one: none after any heapdump answer other than 401, 403 or 404 (a confirmed dump, another 2xx, a 5xx, a redirect, an error or a timeout), and aggressive is for sites you run. An endpoint counts only when its answer has that endpoint's shape, never on status alone, so a single-page app's catch-all page is not a finding. Spring Boot 1.x's gzipped heap dump counts when it opens with the gzip magic and is sent as a .hprof or .hprof.gz attachment (evidence hprof-gzip). An open env, configprops or heapdump is critical, the index and mappings are warnings, and so is a heapdump only listed by an open index. Each request has a 10 second timeout, a timeout ends that base, and the rule stops after 30 seconds; a 429, a bot challenge, a 5xx, a path robots.txt disallowed, or a run that did not reach every path is reported as not checked, never as a pass. The heapdump has its own checks (exposed-actuator-heapdump, -heapdump-closed, -heapdump-not-checked): only an aggressive run that gets a 401, 403 or 404 for it under every base clears it, so an active run never clears a heap dump an earlier aggressive run found. Findings name the endpoint and the kind of evidence, never a value from the response. A path that names a state-changing action, or a POST, PUT, PATCH or DELETE call site, is never probed, and on a passive run the rule reports not checked. See security/exposed-actuator.
  • security/client-side-authz, a probing rule for the classic vibe-coded authorization bug: the frontend hides the admin screens behind a role check, but the admin API checks nothing (squirrelscan/repo#1966). It finds admin and other privileged API endpoints your own pages and scripts reference (a path word such as admin, staff or internal), sends each one GET with no cookies, no credentials and no query string, and fails an endpoint whose 2xx answer is JSON that carries data, naming the field names and record count, never a value. A 401 or 403, a redirect, an HTML page, an empty or health-shaped answer, or a 200 that says unauthorized is not flagged. It runs only when probing is active or aggressive, probes only the audited origin, at most eight endpoints, and never a path that names an action or a URL only called with a write method. An unauthenticated probe cannot show cross-tenant access, and the finding says so. See security/client-side-authz. That makes 331 rules, 27 of them in Security.
  • squirrel audit --bypass-cache (or [crawler] no_cache = true) sends Cache-Control: no-cache and Pragma: no-cache on page fetches (squirrelscan/squirrelscan#2032), so an audit started right after a deploy makes a CDN revalidate, and revalidates pages cached by an earlier audit with a conditional request instead of reusing them on max-age or stale-while-revalidate. It is off by default. When it is on, the report lists pages whose response still looks cached (a positive Age, or a HIT or STALE X-Cache, CF-Cache-Status or X-Vercel-Cache) as cacheFreshness in JSON, and a line in the text and markdown reports. A caller that rendered pages without the render cache can have the report say so with the freshRender reconstruct option.
  • @squirrelscan/core-contracts RUN_PHASES now includes publishing, the single phase a hosted run shows while its report is assembled and published (squirrelscan/squirrelscan#2077). visiblePhaseOf maps the raw publish stage names (complete, editor_summary, publish-init, publish-chunk, publish-finalize, publish-finalizing) to it.
  • squirrel evidence export|verify|replay: a versioned, replayable audit evidence bundle (raw and redacted shareable) and [storage] evidence_archive_dir, which stops retention retiring an audit that has no verified bundle (repo#2301).
  • A local report of a very large crawl (over 250,000 stored checks) is now a sampled body: exact score and totals, per-rule check lists folded to a cap, no per-page check lists, and a top-level reportBody field in the JSON report saying what was left out (JSON schema version 1.23). --full-report builds the whole body (repo#2352).
  • Developers: a shared main-text extractor (squirrelscan/repo#2359). extractMainText(doc, { place }) in @squirrelscan/parser turns a parsed page into { text, markdown, proseText, blocks[] }: headings, paragraphs, lists, tables, quotes and code as markdown, each block labelled with the region and section it sat in and the classifier node that said so. With a placer (extractPageMainText in @squirrelscan/page-classifier wires in the page classifier) it keeps main_content and content_block regions and drops headers, navigation, footers, sidebars, dialogs, search, forms and article chrome, plus ad, cookie consent, chat widget, newsletter signup, social share and promo banner sections; where the classifier abstains, landmarks and conservative class and id tokens decide. Code blocks are marked and left out of proseText. It is not a rule and nothing calls it yet, so no rule output changes (getMainContent is untouched). bun apps/cli/scripts/bench-main-text.ts scores it against a gold set (--gold <dir> for your own) and reports the time it adds.
  • Reports rank findings by cross-site prevalence (squirrelscan/repo#2491): within a severity, a rule that fires on more than 70% of audited sites sorts below distinctive ones and carries a "Ranked lower" line in the markdown, LLM and HTML reports. The figures come from packages/report/src/rule-prevalence.json, refreshed from a corpus of JSON reports by apps/cli/scripts/rule-prevalence.ts; the file ships empty, so ordering is unchanged until a corpus run lands. No rule, severity or score changes.
  • links/broken-links details.notFetched now counts action and token links (action) apart from targets outside the crawl scope or disallowed by robots.txt (excluded), replacing the single skipped count (squirrelscan/repo#3017). The counts still add up to unfetchedTargets, and the Not checked hint says an action-only gap is out of reach of any audit.
  • The JSON report now has meta.discovered: how many distinct URLs the crawl knew of, fetched or not, as total, sitemap (queued from the sitemap) and links (everything else: followed links, the start URL, or carried from an earlier audit). Read it with squirrel report --format json or an audit run with --format json (JSON schema version 1.24). (squirrelscan/repo#3079)
  • Accessibility rules now declare the WCAG success criteria they map to, each with its level and WCAG version; the ten whose checks test no success criterion (such as a11y/heading-order and a11y/landmark-one-main) say why instead, and a11y/duplicate-id-aria maps to 4.1.2 Name, Role, Value. JSON issues carry wcag and wcagLevel, the report has a top-level wcagByLevel, and the HTML, Markdown and text reports show the criterion and level. squirrel audit --wcag-level A|AA|AAA (or [rules] wcag_level) only runs the accessibility rules mapped to a criterion at or below that level (AA keeps Level A and AA). Rules with no WCAG criterion are not run under it, at any level including AAA. Such a run is marked as a partial audit that names the level (meta.partialAudit.wcagLevel in JSON, and a partial-audit line in every format including HTML), is not auto-published (like a --rule-include run, only an explicit --publish sends it), and an --rule-include that selects only rules the level leaves out is refused before the crawl. These automated checks cover part of WCAG; passing them does not mean a site conforms. a11y/touch-targets is now tagged 2.5.5 Target Size (Enhanced), Level AAA, and its advice says so. JSON schema version is 1.27. Each accessibility rule page lists its criteria, or why it has none, generated from the rule metadata (squirrelscan/repo#3093).
  • White-label report branding takes an optional brandColor (#rrggbb) (squirrelscan/repo#3130). It applies only to white-label reports. The HTML report accents its header and .md/.json links with it, using black or white text, whichever contrasts more. It applies only when that text reaches 4.5:1 on the colour and the colour reaches 3:1 against white; otherwise the default theme is unchanged. The XML root gets a brand-color attribute; Markdown and text are unchanged. @squirrelscan/core-contracts exports validateBrandColor, pickTextOn and contrastRatio.
  • perf/bad-caching now judges the caching headers of same-site CSS, JavaScript, font and image responses as their own bad-caching-subresources finding (a sample of up to 50, set by max_subresources), separate from the document checks (repo#4492).
  • New rule a11y/reduced-motion: warns when a page has motion with no prefers-reduced-motion handling (WCAG 2.3.3 Animation from Interactions, Level AAA; 2.2.2 Pause, Stop, Hide is not declared because the rule does not look for a pause control). Motion is a looping CSS animation (animation or -webkit-animation set to infinite) in an inline style attribute or in a rule whose selector matches an element on the page, scroll-behavior: smooth in a rule whose selector matches (html, :root, body and * always do), a <video> with both autoplay and loop, <marquee>, an auto-advancing carousel, or a motion library (AOS, GSAP ScrollTrigger, Lottie, animate.css or WOW.js, parallax, Locomotive Scroll, Lenis) named in a script, stylesheet or preload URL path. Handling is an @media (prefers-reduced-motion …) rule in CSS, a script that calls matchMedia with the query, or a reduced-motion media attribute on a stylesheet; the words in body text or a comment do not count. Inline styles, same-site stylesheets and same-site scripts are read; when a linked stylesheet could not be read the page gets a skipped reduced-motion-unverified check, never a pass. Third-party stylesheets are not read, except that a sheet from a font-only host (Google Fonts, Bunny Fonts, Adobe Fonts, Fontshare) holds only @font-face rules and is ignored. Pages with no detected motion and no unread stylesheet gain a reduced-motion pass. Findings never show a stylesheet's credentials, query string or fragment. The first audit after the upgrade runs page rules fresh (rule cache prc-9). (#711)
  • Every Markdown report of an audit that ran has an accessibility checklist, before the report detail section: one row per WCAG 2.2 Level A and AA success criterion (55, without the obsolete 4.1.1) with what the rules found for it on this audit (rules found an issue, rules found no issue, rules did not run, or no rule covers it) and which rules ran, did not run and found something, read from the criteria each rule declares. A rule that ran and found nothing to check counts as run. Every row still needs a manual check, and the checklist says that automated checks cover only part of WCAG and passing them is not conformance. squirrel report --wcag-checklist [markdown|csv] prints only the checklist, as Markdown (the default) or as CSV with empty reviewer columns. A report read with --input does not carry passing checks, so the checklist says on stderr that a rule with no finding reads as not run. A failed or blocked audit ran no rules, so it gets no checklist: --wcag-checklist on one says so on stderr and exits 1. images/alt-text now cites 1.1.1 Non-text Content (Level A), so the checklist counts it on that row, its JSON issue carries wcag and wcagLevel and wcagByLevel lists it (the JSON schema stays 1.27), and its rule page lists the criterion. --wcag-level still filters accessibility rules only, so images/alt-text runs at every level. The WCAG criteria table in @squirrelscan/core-contracts/wcag now lists every Level A and AA criterion. See accessibility coverage (squirrelscan/repo#3094).

Changed

  • A run narrowed by --rule-include or --rule-exclude is marked partial in every format and in its publish payload: meta.partialAudit in json (schema 1.25, with the included categories) and the same marker in text, markdown, xml and llm. The marker is saved with the audit, so squirrel report and squirrel report --publish carry it too, and a Pro rule you excluded is no longer listed as a locked rule. A filtered run is still auto-published only with an explicit --publish (squirrelscan/repo#1082).
  • content/dev-leakage now finds a tier label at any position left of the audited domain (app.dev.example.com, www.staging.example.com), not only the leftmost, and the vocabulary adds qa, uat, preprod, local, int, demo and sandbox to staging, dev and test. Labels match whole (notdev.example.com stays clean). demo and sandbox only ever warn, since they are often real production surfaces. Auditing a host with any of these labels skips the rule as a non-production origin (squirrelscan/repo#1858).
  • The scan line says how many crawled pages returned errors: Scan: 4 pages crawled (2 returned errors) .... The LLM and XML <scan-scope> gain an error-pages attribute (squirrelscan/repo#2404).
  • Local external-link requests use browser-like headers. Bot blocks, timeouts and Cloudflare or Akamai edge errors are informational and do not affect the score; a TLS failure is confirmed broken. The console and LLM reports explain possible false positives and point to Perch. See how external links are checked (squirrelscan/repo#2747).
  • Toggling the page classifier on, off or to a different model no longer re-fetches pages reused from the crawl cache (squirrelscan/repo#3369). A reused page that is stale only because of the classifier is re-read from its stored HTML (with the new classifier, or none), so unchanged pages cost no request. A page with no stored body is still fetched. packages/crawler/scripts/reclassify-bench.ts measures it.
  • The published entity map now summarizes structurally identical entities into counted groups (groups) instead of sampling them, so a large catalogue states the true size of each shape. Rule verdicts still read the full map (#407).
  • perf/bad-caching reads response header names in any case (ETag, Cache-Control, Last-Modified), and a page whose stored record has headers.recorded: false is no longer counted as missing a header. It shows as not checked with the reason headers-not-recorded (squirrelscan/repo#4641).

Fixed

  • core/meta-description no longer reports a too-short description on a page written mostly in Han, kana, Hangul or Thai. The length floor is calibrated for Latin text, so that check is skipped there, and the skip message says why. Too-long and missing descriptions are still reported (#1866).
  • Partial re-audits score carried findings in the same check units as a full audit, so a re-audit's score no longer drifts from a full audit of the same site (squirrelscan/repo#3360).
  • The GraphQL probes (security/graphql-introspection and security/graphql-get-mutations) now read robots.txt as the crawl does (squirrelscan/repo#3390). With respect_robots on, a path that the crawl's own group (robots_user_agent, default SquirrelScan) disallows is no longer probed below aggressive, and a site with no such group gets its User-agent: * rules, as before. Previously the probes read only the * group, so a User-agent: SquirrelScan group that disallowed /graphql was not honoured.
  • An incremental audit that refetches an unchanged page now reports the headers that response carried, not the ones stored with the page (squirrelscan/repo#4765). A site that dropped an X-Robots-Tag: noindex, or added HSTS, without changing its HTML kept the old header in every later audit, and core/robots-meta reported a noindex the site no longer sent. A 200 whose body matched and a reused render (the conditional-render gate's probe) now replace the stored headers; an origin's own 304 updates the headers it resends and keeps the ones it omits, as HTTP caches do.
  • The installer now says why it could not fetch the release manifest (squirrelscan/repo#1595). A failure names each endpoint tried with its HTTP status or curl exit code, tells "cannot reach the network" apart from "the endpoint returned something unusable", and points at the releases page for a manual download. A 200 response with no version in it no longer ends the script silently.

Update to this version

squirrelscan will auto-update, or run this command to update now.

Downloads

macOS

Your platform
Apple Silicon (M-series chip)
squirrel-0.0.184-darwin-arm64, 105.2 MB
Download
Intel
squirrel-0.0.184-darwin-x64, 113.8 MB
Download

Linux

x64 (glibc)
squirrel-0.0.184-linux-x64, 124.1 MB
Download
ARM64 (glibc)
squirrel-0.0.184-linux-arm64, 124.2 MB
Download
x64 (musl/Alpine)
squirrel-0.0.184-linux-x64-musl, 118.0 MB
Download
ARM64 (musl/Alpine)
squirrel-0.0.184-linux-arm64-musl, 116.9 MB
Download

Windows

x64 (Intel/AMD)
squirrel-0.0.184-windows-x64.exe, 130.0 MB
Download

View this release on GitHub

Audit your site in one command

SEO, performance, security, accessibility and agent experience issues, with exact fixes for your coding agent.

Or install it yourself

No account needed for the CLI. Cloud audits include free monthly credits.