Repository navigation
Releases
wardlatch 0.1.19
Four branches in one release: the coverage protocol (PR #37), the identity-equivalence pattern (VU#604762), no SAFE
from a stale list, and the never-a-false-SAFE suite, with what the suite and an attacker's pass found fixed. From
this release the workflow publishes to npm itself, through npm's trusted publishing, and each release carries a
signed SBOM.
Upgrading
Nothing to do: npx wardlatch@latest picks it up. A few rows that read SAFE before now read UNKNOWN or REVIEW, on
purpose: a Homebrew Deno whose keg folder and program name different versions, a program that names two versions, a
cut lockfile or dpkg status file, and a Mac at Apple's newest release once Wardlatch's list of releases is more than
30 days old. Each says why.
Added
- Starlette before 1.0.1, CVE-2026-48710 (
starlette-2026-06, CISA KEV 2 Sep 2026, so "exploited"). A malformed
Host header makesrequest.url.pathdiffer from the path Starlette routed, so a check that allows or denies by
request.urlcan be bypassed. Python MCP servers run on Starlette: the MCP Python SDK requires it. Read from
starlette-<version>.dist-infoin site-packages, and, new, from the Python lockfiles and pin files under the folders
searched:uv.lock,poetry.lock,pylock.toml,Pipfile.lockandrequirements*.txt, as text, nothing run.
Below 1.0.1 is EXPOSED. A pin that is a range, a URL or no version is UNKNOWN, and so is a file cut short, a line the
reader can't read, or an include from another folder. It was declined until now as a developer package (D3). - MCP Python SDK and FastMCP (
mcp-python-2026-10): CVE-2026-116684 (a token for another resource server
accepted), CVE-2026-116685 (credentials sent across a cross-origin redirect) and FastMCP's GHSA-vqqp-x5qm-5qp4 (a
client secret sent to an authorization server an MCP server advertises). Below each fix is EXPOSED (mcp 1.8.0 to
1.29.1 and 2.0.0 to 2.1.1; fastmcp and fastmcp-slim 4.0.0 to 4.1.1). At or past it, the SDK and FastMCP are REVIEW,
never SAFE. The vendors say upgrading alone changes nothing: the fix is a setting off by default, and the row names
it (validate_token_resource=True;issuer=). Not in CISA's catalog. - Xcode before 27.2 (
xcode-2026-10): CoreDevice's symlink file read (CVE-2026-112605) and an Instruments
authorization issue (CVE-2026-112602), and Xcode 27's two sandbox escapes (CVE-2026-86948, CVE-2026-86949). Read
from theversion.plistof everyXcode*.appin /Applications and ~/Applications, matched as a Mac's disk matches
names, and from the Xcodexcode-selectpoints to, each install once. A beta (Xcode-beta.app) is REVIEW at the
fixed number. Not in CISA's catalog. - Docker Desktop before 4.97.0, CVE-2026-114409 (
docker-desktop-2026-10): third-party API keys and OAuth
tokens written in cleartext to its logs and diagnostics bundles (and, before 4.94.0, the Docker Hub token,
CVE-2026-123082). Below 4.97.0 is EXPOSED. At or past it, Docker says logs containing credentials can remain after
upgrading. Wardlatch lists Docker Desktop's log folder by file name and date only, never a file's contents. A file
older than 4.97.0's release, or older than this copy of Docker.app, makes the row REVIEW, naming the folder and
telling you to delete the logs and rotate the tokens. A folder that can't be listed is UNKNOWN. Not in CISA's catalog. - Fix age on every EXPOSED line: "fix public for N days (since YYYY-MM-DD)", and
fixAgeDays(withfixAge) on
each EXPOSED row in--json. Attackers can now turn a public fix into an exploit within hours, so the line says how
long the fix has been out. Every incident file carriesfixPublished, the advisory's date with its source. Where
there is no fixed release (a malicious release, a leaked key, a compromised authority) it carries no date and the
reason, and the line says "fix date unknown". "Today" is the same as every other date check's (lib/clock.mjs). - A daily coverage sweep in CI (
.github/workflows/coverage-sweep.yml,scripts/coverage/gap-issues.mjs). Every
day, and by hand with no inputs, it reads the sources and runs the coverage check. It opens or updates one issue per
gap ("Coverage gap: ()", labelcoverage-gap) with the job's own token,contents: readand
issues: writeonly. It never commits, pushes or opens a pull request. The metric is "fix public → named":
baseline 91 days, target under 24 hours (docs/COVERAGE-PROTOCOL.md). - Bridge advice for a bundled Brotli (docs/BRIDGE-BROTLI.md). When Pillow, opencv-python or
opencv-contrib-python carries a Brotli decoder without the fix and has shipped no fixed release, the EXPOSED line's fix
adds "While no fixed release exists:" and points to an opt-in, reversible swap for Homebrew's fixed Brotli, on a
Mac. The page gives exact commands, a backup first, a check and an undo, and warns that a reinstall reverts the swap.
Text only: Wardlatch changes nothing. The commands were run on copies of the two packages'.dylibsfolders, never
on the installed ones, and the undo gave back the original files byte for byte. - Class first: the coverage report and the gap issues list identity- and equivalence-class gaps before the others.
- Two new patterns, P8 and P9 (docs/PATTERNS.md). P8: a self-erasing compromise (a commit
reverted, its run and logs deleted) is not "never ran", so an incident of that kind must read history and say UNKNOWN
when history is absent. P9: the fixed version is not the fixed state (a fix that is a setting, or residue an update
leaves). Two new incident keys enforce P9:fixNeedsConfig(REVIEW at or past the fix) andresidueWhenFixed
(what an earlier version left, by name and date). - macOS Audio flaw CVE-2026-65339 (
apple-audio-2026-08, a new kind of check,apple-cve,lib/macos.mjs
appleCveRow). Apple's "Audio" entry, "An app may be able to leak sensitive user information", fixed in Tahoe
26.6.2 (17 Aug 2026) and Sequoia 15.8 (14 Sep 2026). Below the fix on its own line is EXPOSED; Sonoma 14 and older,
a beta, an unread build or an unplaceable version is UNKNOWN; off a Mac the row says iPhones and iPads are not
checked. CISA's catalog (2026.11.13) does not list it. Tests:test/apple-audio-2026-08.test.mjs. - CERT/CC VU#604762, a Mac's disk opening one file for two spellings of its name (16 Nov 2026): Deno before 2.7.14
(deno-2026-06, CVE-2026-49401,--deny-*rules), Terraform built with go-slug before 0.18.3 (terraform-2026-08,
CVE-2026-14978,.terraformignore; read from the go-slug line Go writes into the program) and setuptools before
83.0.0 (setuptools-2026-07, CVE-2026-59890,MANIFEST.in). EXPOSED below the fix on a Mac; on Linux, where the
vendors say the flaws are not, SAFE with that reason (Deno, Terraform) or no row (setuptools). Nothing is run. Two
new keys for version files:scan(a program's bytes read in pieces) andonlyOn(the systems the flaw is on).
Every Deno is read from its program, never from a folder's name, which need not match it: mise's and asdf's
(installs/deno/<version>/bin/deno), the official installer's~/.deno/bin/denoand Homebrew's keg
(Cellar/deno/<version>/bin/deno, reached byopt/denoandbin/deno). A program that can't be read, or carries no
version, is UNKNOWN, and so is a Homebrew keg whose folder and program name different versions. The version is the
user agent Deno compiles in,Deno/<version>+<commit>, measured on Deno's own 2.7.13 and 2.7.14 macOS arm64 release
builds (16 Nov 2026; COVERAGE-PROTOCOL D8), which read 2.7.13 EXPOSED and 2.7.14 SAFE.
Tests:test/equivalence-2026-11.test.mjs. - ConnectWise ScreenConnect client, CVE-2026-84869 (
screenconnect-2026-09, CISA KEV 11 Sep 2026). PR #37's
coverage list declined it as a "self-hosted server": the flaw is in the client, and ConnectWise says servers are not
affected. Reads the host client's and the access agent'sInfo.pliston a Mac: below 26.6.5 is EXPOSED; at or past
it REVIEW (no real bundle read yet); a Linux agent folder UNKNOWN; an agent named for the attacker instance in
Huntress's report EXPOSED.
Fixed
- Eight false SAFEs found by an attacker's pass (
test/never-false-safe/attack.test.mjs, A1 to A9, each with a
planted mutation). A program naming two versions that would be judged differently (a fixed one's text ahead of its
own) is UNKNOWN, not the first one's verdict, and a scanned program not read to its end is UNKNOWN (A1). Terraform
built with go-slug replaced (a=>line) is UNKNOWN (A2). Homebrew'sbin/denoandbin/git, the copies a shell
runs, are read besideopt/(A3). A PATH entry written~/bin, which bash and/bin/shsearch, is read as they read
it instead of being dropped (A4). On an updated Mac, a shell start-up file that namesSSL_CERT_FILE,SSL_CERT_DIR
orCURL_CA_BUNDLEkeeps the built-in curl and python3 from being judged by/etc/ssl/cert.pem(A5). PYTHONPATH's
folders are searched as site-packages is (A6). Lockfiles: classic yarn'snpm:alias read from its spec, a listed
name installed from another package ("react-server-dom-webpack": "npm:@acme/fork@1.0.0") is UNKNOWN in npm and yarn
lockfiles, and a pnpm lockfile re-indented or written inline is UNKNOWN, not "pins nothing" (A7, A8). One is left
as a todo for the owner: the same alias in a pnpm lockfile'simporterssection (A7t). The suite's clock preload now stands in for every filelib/clock.mjsreads, so H11c no longer turns REVIEW by
itself once this Mac's own update files are a month newer than the list (A9). - **A Homebre...
wardlatch 0.1.18
Nine new checks, a fix to the Chrome check, a corrected pip line, and a new mode, --reach. Each new check is a bullet
in the list below and an entry under "Added"; the corrections are under "Fixed". Every check reads files only and runs
nothing it finds.
- React Server Components, CVE-2026-126187 (
react-rsc-2026-11). React's advisory names three npm packages,
react-server-dom-webpack,react-server-dom-parcelandreact-server-dom-turbopack, in versions 19.0.0 through
19.0.8, 19.1.0 through 19.1.9, 19.2.0 through 19.2.8 and 19.3.0, fixed in 19.0.9, 19.1.10, 19.2.9 and 19.3.1 (released
14 Oct 2026; the CVE record was published on 10 Nov). It does not namereactorreact-dom. Wardlatch reads your
projects' lockfiles and names each one that pins an affected version. Next.js carries its own copies of two of these
packages, which no lockfile shows, and has its own advisory (GHSA-673q-mjm6-r5xr, no CVE number):next16.0.0
through 16.4.0, fixed in 16.4.1, is reported under that advisory's name. Neither advisory names the other. - Fortinet FortiClient for Mac, CVE-2026-121529 (
forticlient-2026-11). Fortinet's advisory FG-IR-26-194 of
10 Nov 2026 lists FortiClientMac 7.4.0 through 7.4.7 as affected, fixed in 7.4.8, and FortiClientMac 8.0 and 7.2 as
not affected. Wardlatch reads the installed version from the app'sInfo.plistand never starts FortiClient. A
FortiClient at or past the fix reads REVIEW, not SAFE, and says why: no real FortiClient bundle has been read by
this check yet, and whether FortiClient can keep running an old version after an update is not known. - The tampered Claude Code extension builds 2.1.312 and 2.1.313 (
claude-code-extension-2026-10), served from
Cursor's VS Code Extension Marketplace on 29 Oct 2026 with an SSH backdoor: Cursor's indicators, read from files
(extension bundles and packages by hash, the attacker's key inauthorized_keys, the dropped LaunchAgent and scripts,
the tunnel key). Uninstalling the extension is not enough, and the fix line says so. - macOS behind Apple's latest security release (
macos-2026-11): Tahoe below 26.7.1 and Sequoia below 15.8.1 lack
Apple's fix for CVE-2026-86950, which CISA lists as exploited; Sonoma, which Apple no longer updates, is REVIEW. A
beta is never SAFE. - Visual Studio Code before 1.145.3 (
vscode-2026-11): Microsoft's 16 advisories of 10 Nov 2026, three of them code
execution through MCP servers. - Claude Code before 2.1.251 on macOS (
claude-code-2026-11, CVE-2026-88968), read from its installs without running
it; old copies left beside the one that runs are REVIEW. - Git before 2.56.1 and its backports (
git-2026-11, CVE-2026-111760 and five more); Apple's git 2.54.0 is REVIEW,
since Apple backports. - GPUStack before 2.3.1 (
gpustack-2026-11): below 2.2.0, CVE-2026-102895, which CISA lists as exploited; up to
2.3.1, a critical code execution GPUStack fixed in 2.3.1. - Brave, Vivaldi and Microsoft Edge are read by
chrome-2026-11too, and a fixed Extended Stable Chrome no longer
reads EXPOSED (under "Fixed").
CISA's catalog (version 2026.11.10, read on 11 and 13 Nov 2026) lists neither React's nor FortiClient's CVE (Fortinet's
page says Known Exploited: No; React's advisory says nothing about exploitation), nor the Claude Code, VS Code, git or
tampered-extension issues; those checks print no CISA line and say "fixed in". It lists macOS's three CVEs and
GPUStack's CVE-2026-102895, and those checks print CISA's dates under their titles.
Sources: React's advisory,
the CVE record for CVE-2026-126187,
Next.js's advisory,
Fortinet's advisory FG-IR-26-194,
the CVE record for CVE-2026-121529,
Cursor's notice,
Apple's security releases,
Microsoft's VS Code advisory GHSA-66hq-jw93-462h,
Anthropic's advisory GHSA-5xjx-9qx6-jcrv,
GitHub's post on the git releases,
GPUStack's advisories GHSA-gj9m-f8ph-pmhr
and GHSA-gjf9-3p3r-9h9c,
Chrome's Extended Stable update of 9 Nov,
CISA's catalog. The tests are
test/react-rsc-0.1.18.test.mjs, test/forticlient-0.1.18.test.mjs, test/claude-code-extension-0.1.18.test.mjs,
test/macos-0.1.18.test.mjs, test/vscode-0.1.18.test.mjs, test/claude-code-0.1.18.test.mjs, test/git-0.1.18.test.mjs,
test/gpustack-0.1.18.test.mjs, test/chrome-0.1.18.test.mjs and test/pip-wording-0.1.18.test.mjs. An independent
review of the new checks (13 Nov 2026) found three false-SAFE or false-reassurance paths before release; each is
fixed, says so in its entry below, and has a test.
Added
react-rsc-2026-11, and a new kind of check: lockfiles (lib/lockfiles.mjs). Under the folders searched (10
levels;node_modulesand.gitare not entered) it readspackage-lock.json,npm-shrinkwrap.json,yarn.lock
(Yarn 1 and Berry) andpnpm-lock.yamlas text. It runs and imports nothing from a project, does not open
node_modules, and asks no registry. A pinned version in an affected range is EXPOSED; one outside them is SAFE,
with the lockfile and the version as its receipt. UNKNOWN, never SAFE: a lockfile that can't be read, is not one, is
cut off, or holds an entry the reader does not understand (the whole file); a canary, experimental or pre-release
build, which the advisories do not name; Bun's lockfiles (bun.lock,bun.lockb), which are named and not read
yet; a folder that can't be opened, a link to a folder, and a folder past the depth limit. A lockfile says what an
install would fetch, not what is innode_modulesor running on a server, and the line under the result says so.forticlient-2026-11. Reads/Applications/FortiClient.app/Contents/Info.plistand the same path under
~/Applications. 7.4.0 through 7.4.7 is EXPOSED. 7.4.8 or later, the 8.0 line included, is REVIEW, not SAFE (the
owner, 12 Nov 2026), with the reason on the line: no real FortiClient bundle has been read by this check yet, and
whether FortiClient can keep running an old version after an update is not known. A REVIEW line does not change the
exit code, as with a fixed Chrome. The 7.2 line, which Fortinet says was never affected, is SAFE. A version older
than 7.2 is UNKNOWN, because Fortinet's table names no such line. A version that can't be read (no version
key, a binary plist, a file that can't be opened) is UNKNOWN. FortiClient for Windows, which the advisory also lists,
is named under what the check does not cover. No FortiClient was installed on the Mac this was written on: the
check was tested on anInfo.plistwritten to Apple's format, not on a real bundle.notNamedBelowin a version rule (lib/binaries.mjs): a version on a release line older than any the source
speaks of is UNKNOWN. Without it such a version read as older than the affected range, SAFE, on the strength of
a table that never mentioned it.reviewFrombesidereviewWhenFixedon a version file (lib/installs.mjs): only a SAFE at or past that version
becomes REVIEW, so a version older than the affected range (FortiClient 7.2) is not told it has the fix. Without the
key, any SAFE does, as before (Chrome, GitLab).--reach, a new mode: what a program started from this terminal could open. For a fixed list of places
(Messages, Mail, Safari and Notes data and the privacy database on a Mac; Chrome, Chromium, Edge, Brave, Vivaldi and
Firefox data folders; the login keychain file;~/.ssh,~/.npmrc, the GitHub CLI's hosts file, Docker's config,
AWS, Google Cloud, Azure, Kubernetes, PyPI, Cargo,.netrcand Git's credential store), Wardlatch opens each for
reading and closes it at once, and printsreachable,blocked,not there,could not tellornot applicable.
It never reads a byte and never lists a folder: one test runs it against a stand-in for the disk that fails on any
call but lstat, realpath, open and close, another runs the command under a spy on every read, and both are shown to be
able to fail. It is not a malware scan and has no signatures. It runs instead of the incident checks, never beside
them (--reachwith--incident,--pathsor--no-containersis refused), prints no SAFE, EXPOSED or UNKNOWN
line, and exits 0 whatever it finds; areachableand acould not tellare REVIEW lines. "No such file" is not
taken for "not there" where the system keeps the file in every account, or where a variable the vendor documents
says the tool keeps it elsewhere; a link that leads outside the home folder is not followed; a FIFO or device is not
opened. Each place has the vendor page that names it, and eachreachableline one fix line with the page it comes
from, inlib/reach/items.json. Off ...
wardlatch 0.1.17
eMudhra revoked eight more CAs on 7 November, for a different reason, and they are now REVIEW lines on a Mac. The
October check reads exactly as it did.
- Eight eMudhra RSA CAs revoked on 7 Nov 2026 (
rsaRevocationinemudhra-2026-10). They areEM DV/OV/EV TLS CA - G2B/G2CandEM OV CS CA - G2B/G2C, all RSA-4096, revoked with reason cessationOfOperation. eMudhra says their key
material was handled in a Python process's memory on the ceremony host, so it cannot demonstrate that the keys stayed
inside the HSM. It also says the keys' integrity remained intact
(Mozilla bug 2083449, comment 82). That is not the
October finding, so Wardlatch reports them as REVIEW, never EXPOSED. The six TLS CAs chain throughemSign Root TLS CA - G1toemSign Root CA - G1, one of the five roots Apple's 9 November update removes from
/etc/ssl/cert.pem. Mozilla and Chrome stop trusting what G1 issues after 16 Nov 2026; a CA file carries no such date. - REVIEW lines on a Mac.
/etc/ssl/cert.pemgets one when it holds G1 or one of the CAs above, on any macOS
release, with what to do on that release. So does a copy of it in/etc/ssl(on a release Apple names) that holds
none of the October roots or keys but holds one of these, such as the Mac fix's own cleaned file: 0.1.16 said nothing
of it. REVIEW lines don't change the exit code. - The note no longer says G1 "changes no result". For
emSign Root CA - C1andemSign ECC Root CA - C3it says
what it said. - The October code-signing CAs,
EM OV CS CA - G3BandG3C, and their two cross-certificates underAAA Certificate Services, are in the record (comment 82, Appendices A and B). 0.1.16 named "two sign code" but carried
only the six TLS keys. They are matched by fingerprint only, because their certificates aren't in this repository.
Their lines are REVIEW, because whether a TLS reader trusts a code-signing CA in its CA file was not measured.
Not covered, and said so in the run's last line: certifi, Node.js and Linux bundles are not read for these keys. All of
them still hold emSign Root CA - G1, and no certifi or Node.js release without it exists yet (11 Nov 2026). The Mac
fix still removes only the October roots and keys, not G1.
The certificates are test/support/fixtures/emudhra-2026-11-g2-cas.pem, eMudhra's list is
test/support/fixtures/emudhra-2026-11-bug2083449-c82.tsv (both described in that folder's README), and the tests are
test/emudhra-g2-0.1.17.test.mjs and new cases in test/apple-update-0.1.16.test.mjs and
test/apple-fix-0.1.14.test.mjs.
Verifying this release
The npm package and the source archive carry a Sigstore attestation bound to the workflow run and commit
that built them. There is no key to import.
npm pack wardlatch@0.1.17
gh attestation verify wardlatch-0.1.17.tgz --repo deimos-hq/wardlatch \
--signer-workflow deimos-hq/wardlatch/.github/workflows/release.yml --source-ref refs/tags/v0.1.17 \
--deny-self-hosted-runners
See docs/RELEASING.md for the offline check with the attached bundle, and for the signed tag.
wardlatch 0.1.16
Two new incidents, CISA's dates under five, and Macs that took Apple's 9 November update now get answers instead of
"could not be tested".
- Brotli, CVE-2026-111497 (
brotli-2026-10). The decoder flaw in Node.js, Bun, Python packages (Brotli, brotlicffi,
Pillow, opencv-python) and Brotli libraries. CISA lists Google's Brotli as known to be exploited (added 27 Oct 2026,
due 10 Nov). Wardlatch reads each decoder's own bytes, because version numbers mislead here (Node.js 26.12.1 is fixed
and reports Brotli 1.2.0). On a Mac it also says what Apple's notes of 9 Nov say of macOS's own copy. - Google Chrome, CVE-2026-125350 (
chrome-2026-11). Fixed in 157.0.8103.50 (.50 or .51 on Mac and Windows) on
9 Nov; Google says an exploit exists in the wild, and CISA added it to its catalog on 10 Nov (due 24 Nov). Wardlatch reads the installed version and never starts Chrome. - CISA's catalog dates under every incident CISA lists as exploited: SSHole, Brotli, PHP, Plex and Chrome.
- After Apple's update (macOS 27.2.1, 26.7.3, 15.8.3): 0.1.14 and 0.1.15 read the built-in curl and python3 as
"could not be tested" on an updated Mac, because those tools now refuse the compromised certificates even when the
withdrawn roots are put back. That refusal is now counted for what it is, Apple's own block.
Sources: CISA's catalog,
Brotli 1.2.1,
Apple's notes for macOS 27.2.1, 26.7.3
and 15.8.3, Apple's eMudhra page,
Chrome's stable update of 9 Nov.
The two cert.pem files measured on one Mac on 10 Nov 2026, before and after its update to 27.2.1, are described in
test/support/fixtures/README.md (the other measurements quoted here are in our inspection records, not in this
repository); the tests are test/brotli-0.1.16.test.mjs, test/brotli.test.mjs,
test/kev.test.mjs, test/apple-update-0.1.16.test.mjs and test/chrome-0.1.16.test.mjs.
Security
One defect in released code that could read SAFE where something was wrong.
- A path with
..after a link was cleaned up as text, so Wardlatch read, ran or hashed another file than the one
your programs use (0.1.0 to 0.1.15). Found on 25 Oct 2026 while building the Brotli check (fixed there in
28a72c2), recorded on 1 Nov as a defect in released code, and shipped only now: past SECURITY.md's 7 days, because
it was not split out of an unfinished branch. Details under Fixed below.
Added
-
CISA KEV lines. An incident whose CVE is in CISA's Known Exploited Vulnerabilities catalog prints one line under
its title, in CISA's own vendor and product words ("CISA KEV: Google Brotli CVE-2026-111497 is known to be exploited ·
added 27 Oct 2026 · due date for US federal agencies 10 Nov 2026 · "), and--jsoncarries the rows with their
links underincidents[].kev. Wardlatch never fetches the catalog:scripts/kev-snapshot.mjscopies the rows from a
whole read of CISA's feed before a release (it refuses a partial one), andtest/kev.test.mjsfails if an incident
file and that copy disagree in either direction. Catalog 2026.11.10: SSHole, Brotli, PHP (two CVEs), Plex (four) and
Chrome are listed; GitLab's CVE-2026-110264 is not. A date that is not a real day, or a due date before the date
added, stops the run. -
chrome-2026-11. Reads Chrome'sInfo.plistin/Applicationsand~/Applicationsand dpkg's record of
google-chrome-stable. Below 157.0.8103.50 is EXPOSED. At or past it is REVIEW, not SAFE: a Chrome left open since
it updated runs the old version until it restarts, and no file says which version is running. Other Chromium
browsers, Chrome's other channels and Chrome from snap, Flatpak or an RPM are named under "not covered". -
macOS's own Brotli. On a Mac the Brotli check gives one row from the macOS version and Apple's notes of 9 Nov
2026, and says that is all it rests on (Apple's copy is inside the system, not a file Wardlatch reads): SAFE on
27.2.1, 26.7.3, 15.8.3 or later; EXPOSED, with "update macOS", before them; could not be tested on macOS 14 or older,
for which Apple names no update. It stands in for no other row: the all-clear for copies on disk is still given. -
Leftover copies of
/etc/ssl/cert.pem. On a release Apple names, each/etc/ssl/cert.pem.*that still holds a
withdrawn root (by key) is a REVIEW line, the Mac fix's own backup included. Apple says not to restore a backup made
before the update, and to delete it once you have updated; Wardlatch adds: don't restore it even if a program stops
connecting after the update. Read, never changed. -
pip's own CA list (
pip/_vendor/certifi/cacert.pem, in the machine's Python folders and in apps' Pythons) is read
by key: a REVIEW line when it still holds a withdrawn root, saying which pip on which Python it is. pip 24.2 and later
on Python 3.10 or later ask the system's trust and use this list only with--use-deprecated=legacy-certs(on one
Mac on 27.2.1, pip 26.1.2 and 26.2.1 given that flag connected to a compromised test host of the emSign family);
older pips, and any pip on Python 3.9 or older, use it for every download (pip's ownindex_command.py). Whether
such a line should be EXPOSED is left open: whether a download is refused depends on that Python's TLS library. -
certifi inside apps' own Pythons (Blender, Unity's Emscripten Python) is read by key: a REVIEW line when it holds
a withdrawn root. Those Pythons' plainssldefault reads no list; their bundled certifi and requests accepted
compromised hosts on one Mac. Whether the apps themselves connect that way was not shown. -
Read by key. Where a CA file is judged as a file (
lib/macos.mjsbundleHolds), each certificate's public key is
compared with the two withdrawn roots and the seven withdrawn keys, so a relabelled, re-issued or cross-signed copy
is caught; a block that can't be read keeps the line at "could not be tested", never SAFE. -
brotli-2026-10: the Brotli decoder flaw (CVE-2026-111497), in Node.js, Bun, Python packages and Brotli
libraries. Built over twelve review rounds (24 Oct to 1 Nov 2026) on a side branch, and landed in 0.1.16. A new kind of check
(bundled,lib/brotli.mjs) that reads the decoder itself instead of a version number, because the version numbers
mislead here: Node.js 26.12.1 is fixed and reports Brotli 1.2.0, Homebrew's node uses Homebrew's Brotli, Debian's
fixed packages keep 1.0.9 and 1.1.0. The fix added an error code, so a fixed decoder's error table holds the string
_ERROR_FORMAT_BLOCK_SWITCH. Others used that error name as the sign of the fix before this check:
OrchestratorInc/agent-orchestrator pull request #7107 looked for the string in a library file, and baealex/BLEX pull
request #580 asked the decoder for it at run time (BrotliDecoderErrorString(-17)), both on 23 October. Wardlatch's is
the same idea, built into the check: an exact, NUL-delimited match plus a control string every decoder has
(_ERROR_FORMAT_BLOCK_LENGTH_1), reading bytes only. Trail of Bits' brotli-checker is another tool that finds bundled
copies; it checks source trees and ELF builds, and Wardlatch points to it.- Brotli 1.0 writes the same table without the prefix (1.0.x's
BrotliDecoderErrorStringhasreturn #NAME;, 1.1.0
onreturn #PREFIX #NAME;; both read in google/brotli's ownc/dec/decode.cat v1.0.9, v1.0.10 and v1.1.0), so
Debian 12's libbrotli1 1.0.9 holdsBLOCK_LENGTH_1, and 1.0.10, the fix on the 1.0 line, addsBLOCK_SWITCH
(code -17, right afterDISTANCE). Those short words could be in any program, so that form is judged only as a
whole table (maintainer, 2026-10-25): the names fromHUFFMAN_SPACEtoDISTANCEin Brotli's order
aroundBLOCK_LENGTH_1, each NUL-delimited (up to 16 NULs between two, as a compiler pads strings it aligns),
either way round (the orderdecode.hlists them, as clang lays them out, or by code), thenBLOCK_SWITCHand
DICTIONARY_NOT_SET(1.0.10),DICTIONARY_NOT_SETat once (1.0.4 to 1.0.9), orCOMPOUND_DICTIONARYand
DICTIONARY_NOT_SET(1.0.3, the first affected release, whosedecode.hlistsDISTANCEthen
COMPOUND_DICTIONARY). - What proves SAFE: the marker exactly as often as the control, in the file (in each build of a universal Mach-O file,
with every Brotli library a file with a table also links SAFE too) or in the Brotli library it links (every
candidate, by its real path, named in the receipt), or a Debian package version its table lists as fixed, for the
very file that package installed (dpkg's md5sums list names it with the same MD5). In Brotli 1.0's form: every copy
ofBLOCK_LENGTH_1in a whole table withBLOCK_SWITCHin its place, andBLOCK_SWITCHnowhere else. - What is EXPOSED: the control without the marker (or fewer markers than controls), in the file or its library; a
whole Brotli 1.0 table withoutBLOCK_SWITCH(a fixed 1.1-form table beside it included), which for a Debian
library hands over to its package version as the 1.1 form does (Debian 12's own 1.0.9-2+b6 reads EXPOSED by both;
its security build 1.0.9-2+deb12u1 addsBLOCK_SWITCHin its place and reads SAFE by its table alone); a Debian
package built from a source version its table lists as without the fix (every rebuild, such as 1.1.0-2+b7, keeps
its source's version), again for the file that package installed. A bundled copy with no fixed release yet
(Pillow, opencv-python) is EXPOSED, and its fix line says no release had a fixed Brotli on 10 November and to...
- Brotli 1.0 writes the same table without the prefix (1.0.x's
wardlatch 0.1.15
The first new incident since 22 October: SSHole, CVE-2026-123158. sshd in OpenSSH before 10.6p2, on Linux with musl
libc and without PAM, accepts a very long password for an account whose password field is * (no password set, as for a
key-only account) when another local account has a $5$ or $6$ hash. Alpine's advisory says 257 to 1024 bytes; the CVE
record says between 256 and 1024 characters. Alpine is such a system. CISA added it to its Known Exploited
Vulnerabilities catalog on 9 Nov 2026, due 12 Nov. Sources:
Alpine's advisory (5 Nov),
the CVE record,
OpenSSH 10.6p2's release notes,
Alpine's secdb for each branch's fixed build, and
CISA's catalog.
This fixes no security defect in Wardlatch, so there is no advisory. test/sshole-0.1.15.test.mjs holds the tests.
Added
sshole-2026-11, Linux only, read-only. It never runs sshd (not evensshd -T) and never connects to a port.- musl or not:
/etc/alpine-release, or musl's loader/lib/ld-musl-*.so.1. A glibc system is SAFE only when
every sshd program found names glibc's loader; one that doesn't may be built against musl, and is UNKNOWN. - The package:
openssh-server, and Alpine's two PAM buildsopenssh-server-pamandopenssh-server-krb5(both
built--with-pam, per aportsmain/openssh/APKBUILD), in
/lib/apk/db/installed, judged against Alpine's fixed builds (edge 10.6_p2-r0, 3.24 10.3_p1-r2 … 3.16 9.0_p1-r6).
Alpine backported the fix, so on 3.24 10.3_p1-r2 is fixed though it is below 10.6p2, and 10.3_p1-r1 is not. At or
above 10.6_p2 a build is fixed anywhere. Below it, only the fix of the branch/etc/alpine-releasenames counts:
revision numbers mean nothing across branches (edge built an unfixed 10.0_p1-r11 in September 2025, the very
version of 3.22's fixed build), so a newer build pinned from another branch is UNKNOWN, and on edge, on 3.15 or
older, or where the release can't be read, anything below 10.6_p2 is not fixed, except a version that is one of
Alpine's fixed builds, which can't be told apart there and is UNKNOWN (so is 3.23's fixed build on 3.24). A version
that can't be read is UNKNOWN. An sshd program apk didn't install, or a link where apk installed one, makes a
fixed package UNKNOWN. Beside an unfixed one it is named as not read, and since it may read a configuration other than
/etc/ssh's,
nothing in/etc/sshcan make the row REVIEW then; only the account reasons can. - A fixed package with an older daemon: Alpine says upgrading doesn't restart a running sshd. A running
sshd
(orsshd.pam,sshd.krb5) that started before its program was installed is UNKNOWN with "restart sshd". The
install time is the file's inode change time, since apk sets each file's modification time from the package. The start
times come from/proc, and are believed only when this process's own start, read the same way, comes out right. - An unfixed package is EXPOSED, unless something Alpine's advisory says the flaw needs was read as missing,
which makes it REVIEW (still to update):PasswordAuthentication nowith noMatchblock setting it again, a PAM
build alone withUsePAM yes, no account with*that the flaw can use, or no$-hash. Any setting in
/etc/conf.d/sshd, a second sshd instance, or a running sshd started with-oor-fon its command line (read
from/proc/<pid>/cmdline) makes the settings "not read", as does a command line or a process list this can't read,
or anIncludeit can't read. A REVIEW line names what wasn't read, as an EXPOSED line does.
/etc/shadowis read only when this user may (usually root), and only counted: no name, field or hash is printed.
/etc/passwdis read for accounts with*and no shadow line. Root counts only withPermitRootLogin yes, as
Alpine says. Accounts from a directory service can't be seen here, so they never make it REVIEW. What couldn't be
read is named on the line. - The fix is Alpine's:
apk upgrade -U -a, then restart sshd, or turn password logins off until then (after checking
that a key login works). Wardlatch doesn't read logs, so it can't say whether anyone got in; it gives Alpine's
advice on where to look. - Not covered: sshd in containers and images.
- musl or not:
Verifying this release
The npm package and the source archive carry a Sigstore attestation bound to the workflow run and commit
that built them. There is no key to import.
npm pack wardlatch@0.1.15
gh attestation verify wardlatch-0.1.15.tgz --repo deimos-hq/wardlatch \
--signer-workflow deimos-hq/wardlatch/.github/workflows/release.yml --source-ref refs/tags/v0.1.15 \
--deny-self-hosted-runners
See docs/RELEASING.md for the offline check with the attached bundle, and for the signed tag.
wardlatch 0.1.14
Apple shipped its fix for /etc/ssl/cert.pem on 9 November 2026, and from that day Wardlatch's words "no macOS release
has fixed it yet" and "run this again after every update" were wrong. Apple says macOS 27.2.1, 26.7.3 and 15.8.3 remove
the two withdrawn roots from the file, "along with three other eMudhra root CA certificates", and that on these releases
the built-in command-line tools stop accepting the 12 compromised certificates
(Apple, support.apple.com/en-us/149555, as published on 9 Nov 2026). This
release says so. No result changes: SAFE, EXPOSED and UNKNOWN still come from what each tool does with the test hosts,
not from a version number. This fixes no security defect, so there is no advisory. test/apple-fix-0.1.14.test.mjs
holds the tests.
Changed
- The built-in curl's and python3's fix texts follow the macOS release. On a Mac, Wardlatch reads
ProductVersion
from/System/Library/CoreServices/SystemVersion.plist(a file read; nothing is run for it).- On 27.2, 26.7.2, 15.8.2 or any earlier release of those three, or when the version can't be read, an EXPOSED
/usr/bin/curlor/usr/bin/python3is told to update macOS, and to use the Mac fix if it can't. - On macOS 14 or older, for which Apple names no update, it is told to update to macOS 15 or later if it can, and
otherwise to use the Mac fix. - On 27.2.1, 26.7.3, 15.8.3 or later, where Apple says the roots are gone, it is told how to look at the file, that
Apple says not to restore a backup made before the update, and how to remove the roots. - The python3 text no longer says "see the curl line", which wasn't there when curl read SAFE.
- On 27.2, 26.7.2, 15.8.2 or any earlier release of those three, or when the version can't be read, an EXPOSED
- A note under the eMudhra rows on a Mac: the macOS version, what Apple says of it, and whether
/etc/ssl/cert.pem
still holdsemSign Root CA - G1,emSign Root CA - C1oremSign ECC Root CA - C3, the three other roots Apple's
update removes. Apple did not say their keys were compromised, so they are a note and change no result. On a release
Apple says removes them, a file that still holds them is not what Apple says the update installs, and the note says so.
A file from which no certificate can be read says that, rather than nothing.--jsoncarries the note asmacos:
{ version, appleFix, source, notes }. - An UNKNOWN for the built-in curl or python3 on 27.2.1, 26.7.3, 15.8.3 or later says why it may be one. Apple says
those releases block the 12 certificates in the built-in tools themselves, which could make the test's forced-roots
control fail too. We haven't measured that on an updated Mac, so the row stays UNKNOWN and says so, and its reason
no longer says the error was not a block. - The words "a macOS update puts Apple's original file back" are now about a Mac that can't update to those
releases. We measured updates to 26.7.2 (27 Oct) and 27.2 (28 Oct) doing it. We haven't measured what an update to
27.2.1, 26.7.3 or 15.8.3 does to a file already changed by hand, and nothing here says. - The Mac fix download has a new digest, for its words only. Its opening comment, the scope lines under every SAFE
and the message for a damaged file said that no macOS release had fixed the file and that every update undoes the fix.
They name Apple's releases now, and the Undo line says Apple's own limit on it: don't undo after updating to 27.2.1,
26.7.3 or 15.8.3. No line of its logic changed, and no verdict or exit code. Its SHA-256 is
0dae0dc901a23aba216211f4ef6002a6d75b232fe479ef79db463d5c1dc9b98c; a copy with SHA-256893f53cb...needn't be replaced. - The Mac fix's test suite could fail on a run that was fine.
test/fix-macos.check.shpiped the output into
grep -q, and underpipefailgrep stopping at its first match can kill the writer, so a case read "printed no line"
while it had printed two. Seen once in four runs under load; the old form failed 7 times in 300 on a 5,000-line output.
It reads a here-string now.
Verifying this release
The npm package and the source archive carry a Sigstore attestation bound to the workflow run and commit
that built them. There is no key to import.
npm pack wardlatch@0.1.14
gh attestation verify wardlatch-0.1.14.tgz --repo deimos-hq/wardlatch \
--signer-workflow deimos-hq/wardlatch/.github/workflows/release.yml --source-ref refs/tags/v0.1.14 \
--deny-self-hosted-runners
See docs/RELEASING.md for the offline check with the attached bundle, and for the signed tag.
wardlatch 0.1.13
We found every defect under Security below ourselves, on 2 November 2026, while checking the advisories for 0.1.11 and
0.1.12. Each could read SAFE where something was wrong. If 0.1.12 or earlier said SAFE, run this version. The advisory is
GHSA-2q92-ggx3-wq82. test/seat11.test.mjs holds the tests; L1 runs on Linux only.
Security
-
A project file one level past the header walk was not read. A folder past a walk's limit that holds only files
makes no row, so a.mcp.jsonin a project with no.git, five levels below a folder searched, readSAFE. The
walk reads the project files one level past its limit now. -
The quick look stopped at a link.
~/app/projas a link to a project with a stealer in its.venvreadSAFE,
reachable or locked. The quick look now goes through a link that leads to a folder, as deep as it goes into a folder.
A locked target is UNKNOWN. On a Mac, EPERM there is named under the result as~/.Trashis, because~/OneDriveas a
link to a folder macOS protects would otherwise end every run in exit 2. A link past a locked folder can't be told
from a link to a file, and the quick look still reports only one named like a venv. -
A venv whose own name isn't valid UTF-8 had no row at the quick look's last level, on Linux. It is UNKNOWN now.
Changed
- The Mac fix download has a new digest, for its words only. Its comments and three of its messages said that a site
loading "through this file" showed what the file trusts. The Mac's curl also falls back to Apple's trust (seat 16,
measured 2 Nov 2026), so they say what was measured now, and the message for a test site still trusted after the fix
names that third cause. No line of its logic changed, and no verdict or exit code. A copy with SHA-256c8bba160...needn't be
replaced.
Verifying this release
The npm package and the source archive carry a Sigstore attestation bound to the workflow run and commit
that built them. There is no key to import.
npm pack wardlatch@0.1.13
gh attestation verify wardlatch-0.1.13.tgz --repo deimos-hq/wardlatch \
--signer-workflow deimos-hq/wardlatch/.github/workflows/release.yml --source-ref refs/tags/v0.1.13 \
--deny-self-hosted-runners
See docs/RELEASING.md for the offline check with the attached bundle, and for the signed tag.
wardlatch 0.1.12
We found every defect under Security below ourselves, on 2 November 2026, in a second read of 0.1.11 made after it
shipped, three reviews of this release's drafts, and a differential against 0.1.11 over about 250,000 generated cases. Each could read SAFE where something was wrong. If 0.1.11 or
earlier said SAFE, run this version. The Mac fix download doesn't change. The advisories SECURITY.md promises for these
defects, and for those in 0.1.11, are not open yet. test/seat6.test.mjs to test/seat10.test.mjs hold the tests. Two of seat 6's, I4-H2 and R11, pass on 0.1.11: they pin fixes already made. There is no test for the Linux
link name.
Security
-
Apple's own Python folders were never searched.
/Library/Python/3.x/site-packages, wheresudo pip3installs, and
the site folders of the Command Line Tools and Xcode are the ones/usr/bin/python3names, and none was on the list.
A stealer in one readSAFE, exit 0. -
A link to a folder the run couldn't reach was dropped with no row. All four walks asked whether a link led to a
folder, and a locked target is neither a folder nor gone. A venv, repository or WordPress site behind one readSAFE,
exit 0, while the same link reachable read UNKNOWN. On Linux a link with a name that isn't valid UTF-8 fell in the same
hole. Below the folders searched it is UNKNOWN now, and so is a link to a file behind a locked folder, which the run
can't tell from one. In the quick look round the home folder only a link named like a venv is reported. -
The quick look missed a venv it could not enter.
~/app/.venvas a link to a locked target, or a folder of mode
0644, readSAFE. It is UNKNOWN now, and so is any folder of that mode there, whatever its name. -
0.1.11 put a name back in charge of a verdict. Its quick look left out
.Trash,Photo Booth Libraryand
*.photoslibraryby name at any depth, so a locked~/Documents/clients.photoslibraryholding a stealer readSAFEand
the same folder calledworkread UNKNOWN. Now only~/.Trash, andPhoto Booth Libraryand any*.photoslibraryin
~/Pictures, are left out, and only on EPERM, which macOS gives there and a sandbox or security tool can give too.
When they are refused they are not searched, and the line under the result names them, says how many items at their
top level went unsearched, and says how to search them: Full Disk Access, or emptying the Trash. macOS answers for a folder's own metadata
while refusing its contents, so a refused folder holding nothing is proven empty and isn't mentioned. That works on APFS;
on another disk the count is left out. -
The temp-folder fix in 0.1.11 stopped short. With
/private/var/folders, or the level below it, refused, the run
readSAFE, exit 0, withTgone from what it looked at. Those levels are UNKNOWN now, as is the per-user folder through
T's row, exceptzz, which holds only system users. Inside a sandbox that refuses them this can read UNKNOWN where nothing is wrong: it is a place not looked at. -
A
sitecustomize.pychanged by one byte readSAFE. The published file wasEXPOSEDand anything else had no row,
by an option that silenced the line. It is a REVIEW line now, and the option is gone. A file of that name inside a
site-packages turns that check's SAFE into REVIEW, exit 0. Homebrew's own is one folder up and isn't read. A
sitecustomize/package, or either compiled with no source, which Python runs the same way, is a REVIEW line too. The
owner's decision, 2 Nov 2026. -
WordPress leftovers read
SAFE.<site>/content/pluginsbeside a stock site is checked now, asbackup/pluginswas,
and so is a folder under any parent that holds a listed plugin (wp-content/upgrade,backup/old). A folder holding a
listed plugin's main file is that plugin, whatever it is called (rank-math-pro, a pasted version number, a change of
case), in the live plugins folder too, where it had only an "added" REVIEW line. A folder with the plugin's name in it,
set off by anything but a letter or digit (plugin copy,plugin 2,old-plugin), and without the main file, has its
indicator searched and is a REVIEW line. So is a folder of the exact name in a parent with nopluginsin its name; it
was not looked at before. A link at a copy's name counts.--pathsnaming a copy, a link to the plugin, or a folder holding
its main file, is checked as the exact name is. -
Private keys in forms LibreSSL loads read
SAFE. Lengths that aren't minimal, however long; a constructed OCTET STRING;
high-tag forms (3f 10,3f 80 10,1f 02), for the outer element and the ones inside; a version other than a single
byte 0 or 1, however written; a short key with no NUL byte. These go through one reader now, which normalises the tag and
the length, and they are found, or UNKNOWN where they can't be read. A file that merely begins0., such as aVERSIONof
0.1.0, read UNKNOWN, exit 2, and no longer does. -
Header forms read
SAFE. In TOML,--headerorhttp_headersin an inline table, a dotted key split by a quote, a
quoted key holding=, a]or a quoted word in a comment, multi-line strings and escapes: a small tokenizer now reads
the file beside the line reader, and a row from either counts, so a file a strict parser gives up on loses nothing. Text
inside a multi-line string can therefore be a REVIEW line. A file with a multi-line string never closed and no header the
line reader can see is UNKNOWN, as broken JSON is. In git config, a
value continued by a backslash on a file's last line with no line end, quotes inside a value (ada"@"example.org), and a
CR or U+2028 inside one. A FIFO at.mcp.jsonin a project with no.git. A project file in a repository's subfolder is
still not read, in this release or earlier.
Fixed
- A PKCS #1 public key in DER read UNKNOWN, exit 2, for an ordinary public key. It stays quiet now, on its whole shape: two
INTEGERs, a modulus and a small exponent. - A
--pathsfolder that is a plugin's now checks its parent once, however many were named. The sitecustomize advice no
longer lands on a Docker image's row, where it told a healthy machine to rotate every secret. Both were in drafts only. test/round6.test.mjsfailed in place of skipping inside a sandbox.test/fix-macos.check.shsaid it is safe on your own Mac, and on a Mac with a network it contacts the advisory's test
host. Its header says so now, and how to keep it offline.- SECURITY.md said this repository is ahead of the download. It stopped being true on 2 Nov 2026.
Verifying this release
The npm package and the source archive carry a Sigstore attestation bound to the workflow run and commit
that built them. There is no key to import.
npm pack wardlatch@0.1.12
gh attestation verify wardlatch-0.1.12.tgz --repo deimos-hq/wardlatch \
--signer-workflow deimos-hq/wardlatch/.github/workflows/release.yml --source-ref refs/tags/v0.1.12 \
--deny-self-hosted-runners
See docs/RELEASING.md for the offline check with the attached bundle, and for the signed tag.
wardlatch 0.1.11
We found every defect under Security below ourselves, in reviews of our own work from 29 October to 2 November 2026,
and each entry ends with the day it was found and the day it was fixed (dates in US Pacific time, as the commits record
them). The fixes to the command ship in this package. The Mac fix download changes only when the site serves the new
script and its digest together (SECURITY.md, step 4), so a copy downloaded before then has none of its fixes. The
advisories SECURITY.md promises for these defects are not open yet.
Security
-
Folders Wardlatch could not open were passed over in four more places, and each read
SAFE. All four are in 0.1.10
too. A folder name that isn't valid UTF-8 read as not there, on Linux, and everything under it was skipped, in every
walker. A folder the quick look round the home folder could not open was skipped with no row, so a stealer in
~/DocumentsreadSAFE, exit 0, on a terminal without access to it; it is UNKNOWN now, except.Trashand the two
Photos libraries, which macOS keeps to itself on every Mac. A settings file the system refused with EPERM was counted
as not used, when the person's own tools may read it; it is UNKNOWN now. This user's own per-user temp folder that
could not be listed was dropped; it is UNKNOWN now. And aPATHfolder that can be entered but not listed hid
php8.1; for a rule with several names it is UNKNOWN now.test/round6.test.mjs. Found and fixed 2 Nov 2026. -
Fourteen more places a real finding read
SAFE, found and fixed 2 Nov 2026. The git config parser now reads a section
and key on one line, a value continued with a backslash, and a;inside quotes. A TOMLargsarray is read as a JSON
one is. A compromised WordPress copy beside a core, under a folder likeplugins.bak, in a folder named for its
version, or atVersion: 3.0.123.0is matched; a single-file plugin added in the window is a REVIEW line. A dangling
image with a known-bad digest isEXPOSED, and an engine whose list can't be read is UNKNOWN. A LiteLLM tag the rule's
text names, with an unlisted digest, is a REVIEW line. A bare repository, apip --targetfolder, and.mcp.jsonin a
project with no.gitare read. A failing program's warning is no longer taken for its version, a venv with only
bin/python3is probed, and a FIFO in place of a certifi bundle is UNKNOWN.test/seat5.test.mjs. -
A failed control could hide the only EXPOSED the Mac fix can print (never released). Round 5 took the date off
curl's 0 and left the 0 inside the control'sif, so when the ordinary website failed for any reason (a DNS block, a
firewall, a timeout, an inspecting proxy) the test site was not read, and a Mac whose curl verified the advisory's
chain readSAFE, exit 0. 0.1.10 reads the test site first and never had this. The test site is now read before the
control on both paths, the 0 isEXPOSEDwhatever the control says, and the control gates the 60 alone, as the
script's comments already claimed.test/fix-macos.check.shgains the case, with the 60 as its control. Found and
fixed 2 Nov 2026 by the four-seat panel. -
A committed private key written with non-minimal DER lengths was skipped with no row.
derKind()reads the common
encodings and LibreSSL reads more; a key rewritten with a five-byte length, or a two-byte version number, was neither
found nor counted as unread, so the repository readSAFE, exit 0. A file that begins as a DER element and is not
read as one is now UNKNOWN, on the working tree and on git's copy of a deleted file, whatever it is called.
test/round6.test.mjs. Found and fixed 2 Nov 2026 by the four-seat panel. -
The containers' date gate failed open. A missing or unreadable
canaryNotAftermeant "not expired", so containers
were probed past the real expiry and every refusal readSAFE. Latent: today's value parses. It now closes, and
test/round6.test.mjspins it. Found and fixed 2 Nov 2026 by the four-seat panel. -
A
--pathsfolder that isn't there said nothing on the checks that don't search it, and exited 0. The README
promised an UNKNOWN line. It is one now, on every check. A folder that is there and plays no part is named under the
all-clear as not searched, and inpathsNotSearchedwith--json, never in the receipt (AUDIT-11). Found 29 Oct
2026, fixed 2 Nov 2026. -
A NUL byte hid a withdrawn root from the Mac fix, and not from the tools the fix is about. macOS's
/bin/bash3.2readcuts a line at its first NUL byte and drops the rest of it, and so doesawk
(measured:printf 'ab\0cd\n'comes back fromreadasab, length 2). LibreSSL does not: it reads the file
withBIO_getsin 253-byte pieces and cuts each piece at its first NUL, so on a line longer than 253 bytes it
reads every piece after the first. So one certificate block could carry base64 thatcurl,git,ruby,perl
andopenssldecode asAAA Certificate ServicesoremSign ECC Root CA - G3while the fix, reading the same
block line by line, saw an innocent throwaway certificate — and saidSAFE, exit 0, "no longer holds the two
roots. Nothing changed", with the root still verifying through that file before and after. No earlier check caught
it: the-----BEGINmarker still only ever started a line, so the round-3 gate matched nothing; the counts agreed;
and every certificate the fix did read had a fingerprint and a key, so the completeness gate passed too. A
line-length check cannot stand in for it, becauseawkmeasures the line it already cut: on such a file
awk 'length($0)>252'reports no line at all.roots()now counts the NUL bytes on bytes, withtr, before it
reads a line, and any at all makes the run UNKNOWN, exit 2, with nothing changed and a line saying why. Every path
that reads the file for withdrawn roots goes throughroots()— the snapshot, the backup, the re-reads from the
file itself, the cleaned file andbuilt()— so one test covers them all. The same refusal makes true a sentence
that was not:strip()rewrote the file withawk, so every line holding a NUL byte came out cut short, while
the header said it "removes only each withdrawn root's own lines … and nothing else in the file" (measured: Apple's
file plus# admin note: keep\0 this whole linewas rewritten with that line cut, under a SAFE that named only the
roots). Apple's own file holds no NUL byte and no line over 231 bytes, so this costs it nothing.
test/fix-macos.check.shgains three cases: a block built so that LibreSSL reads the withdrawn root and a line
reader does not, for each root, each shown first to be read that way by LibreSSL and to verify the root through the
file (otherwise the case would measure nothing); a NUL byte outside any block, in a file that still holds both
roots, so it is about a run that did rewrite the file; and the control that a file with no NUL byte is still
fixed, without which a change that refused everything would pass. Found 31 Oct 2026, fixed 1 Nov 2026 (73038cb). -
A compromised eMudhra intermediate, or
emSign Root TLS CA - G3, left in the bundle read SAFE. The incident is
a compromise of eight CA keys, six of them TLS, in 12 certificates. Removing the two withdrawn roots clears the
compromised intermediates only for a reader that insists on a chain ending at a self-signed root in the file.
/usr/bin/curl8.7.1 does not: like curl since 7.68 on OpenSSL-style backends it takes every CA certificate in a
--cacertfile as a trust anchor, and Apple'sgitgoes through the same libcurl, as does the built-inperl. So
acert.pemholding one of those intermediates kept trusting everything it had issued after both roots were gone,
and the fix — which looked only for the two roots — saidSAFE, exit 0, nothing changed.emSign Root TLS CA - G3,
which issued six of the 12, is worse again: besides the cross-signed certificate the advisory's hosts serve, it
exists as a self-signed certificate, which is a full trust anchor for every reader,python3and theopenssl
CLI included, with no partial chain needed. Measured on the branch as it stood: all 14 of those certificates read
SAFE, exit 0, unchanged, each still readable out of the file by LibreSSL afterwards; and the advisory'sov-g3c
leaf verified through the cleaned bundle plus its intermediate underopenssl verify -partial_chain, and failed
through the cleaned bundle alone. The fix now carries seven more keys and removes any certificate holding one, by
key, in the same way and with the same machinery as the roots' keys: the six compromised TLS intermediate keys
(which cover all 12 certificates, because each intermediate was issued twice, underemSign Root TLS CA - G3and
underAAA Certificate Services) andemSign Root TLS CA - G3's own key, which covers both of its certificates at
once. A bundle too small to rebuild that holds one of them is UNKNOWN with the key named, not "nothing to remove".
lib/incidents/emudhra-2026-10.jsoncarries the seven aswithdrawnKeys, apart fromdistrustedRoots, which
drives the probes' forced-roots control and has one PEM file and one canary group per entry. Each key value was
derived from its certificates three times — LibreSSL 3.3.6, OpenSSL 4.0.3 and node's owncrypto— and refused if
it came out as the SHA-256 of no input (e3b0c442…).test/fix-macos.check.shgains one case per certificate, 14
of them, each with LibreSSL as the oracle that the key was in the file before the run and gone after it, counted by
the SHA-256 of the SubjectPublicKeyInfo rather than by a name; plus the control that a bundle holding none of them
is left alone, and the small-bundle case.test/incidents.test.mjsgains the data's own checks, including that
none of the seven is in the macOS bundle Apple ships —...
wardlatch 0.1.10
Security
- The Mac fix (
fix/fix-macos.sh) could say SAFE for a file that still trusted the leaked eMudhra keys. It found
the two roots by their=== …/CN=name lines. Some copies of/etc/ssl/cert.pem(on some managed Macs) have no
name lines, so it found neither root and removed nothing. If the advisory's test site then answered, it printed
EXPOSED (exit 1) and changed nothing; if the site didn't answer, it printed "SAFE … no longer holds the two roots"
(exit 0) for a file that still trusted the keys. It now finds the two roots by their SHA-256 fingerprints, removes
them whether or not the file has name lines (on Apple's file, byte for byte the advisory's removal, as before), and
refuses a file with a certificate it can't read (UNKNOWN, exit 2, nothing changed). It also refuses a file holding
blocks curl trusts but the fix doesn't read (BEGIN TRUSTED CERTIFICATE,BEGIN X509 CERTIFICATE): an emSign root
appended in that form would otherwise have read SAFE with the test site unreachable. It works on a snapshot of the
file and won't replace it if it changed meanwhile, it refuses a file it can't read, and its re-test ignores
~/.curlrc(curl -q).test/fix-macos.check.shgains the cases, counts the roots by fingerprint, and checks
that each run prints nothing on stderr and leaves no temporary files; the new cases fail on the old script. The
download on www.deimoscore.com changes with it (SHA-25650e1e41d…21b1). The advisory's own one-line command has
the same blind spot: on a file with no name lines it removes nothing. - The Mac fix now says that a macOS update puts Apple's original file back (macOS 26.7.2 did, on 27 Oct 2026), so it
should be run again after every update; its last line saysnpx wardlatch@latest, and it names the tools the file
covers (the built-in curl, git, python3, openssl, perl and ruby) and that Apple'spip3has its own copy.
Fixed
- The workflow review read SAFE without having read every workflow in the window, in two ways (0.1.0 to 0.1.9).
Both came from giving git the window's start withgit log --since.- The start was a bare date. The
github-actions-2026-10incident gave the window's start as2026-09-01, passed
to git as it was, and git reads a bare date as that day at the current time of day, in the zone git runs in: up
to 0.1.8 the zone of theTZWardlatch was started with (the machine's whenTZwas unset), which git got with
the rest of the caller's environment, and in 0.1.9, which gives git noTZ, the machine's. So the window began on
1 Sep at the hour the check ran: a workflow committed on 31 Aug, or on 1 Sep before that hour, was not read, and
with nothing else found the review said "no workflow files added since 2026-09-01", SAFE, exit code 0. The sources
say to check from 31 Aug (The Hacker News and StepSecurity, with no time zone), and GitGuardian reports a commit
through the same compromised account at 00:30 UTC on 30 Aug. No test had a commit near the start of the window. - git stopped reading at a back-dated commit. Given
--since, git stops following a line of history at the first
commit dated before the start (its walk drops that commit before queueing its parents; only--since-as-filter,
git 2.37 and later, doesn't). Whoever makes a commit sets its date, so one commit dated before the window, on top
of a branch, hid every workflow added beneath it, however those were dated: on 0.1.9's code with git 2.54, a
workflow committed on 10 Oct with a commit dated 1 Jan 2025 on top read SAFE, exit code 0; one beneath a
back-dated commit with another workflow above it was left out; and a workflow on a side branch behind a
back-dated commit, merged inside the window and the branch deleted, read SAFE. 0.1.9's known limits named this
only for a later commit made on a machine with a wrong clock. - Also found while fixing this, each checked on 0.1.9's code with git 2.54. A commit whose committer date git can't
read, on top of a workflow, did the same as a back-dated one (git takes it as 1970): SAFE. A workflow added on a
side branch and removed again before the merge, all inside the window, read SAFE: given a path,git logfollows
only the parent a merge's files match and passed the side branch by. In a shallow clone cut inside the window, a
workflow added and removed again below the cut was never seen (SAFE), and a workflow still there at the cut was
listed as added by the cut commit, at that commit's date. Not a miss in 0.1.9: a replace ref (refs/replace/)
couldn't put another commit in place of a workflow's commit (GIT_NO_REPLACE_OBJECTS=1, since 0.1.7; the commit
the ref points to is read as one more commit), and a.git/info/graftsfile wasn't read
(GIT_GRAFT_FILE=/dev/null, in 0.1.9); both still hold. - The window now starts at 30 Aug 2026 00:00 UTC (
"since": "2026-08-30T00:00:00Z", with its sources in
sinceSource). The review takes only an explicit UTC instant (YYYY-MM-DDTHH:MM:SSZ); anything else is
UNKNOWN, never SAFE. - git gets no date at all.
git log --all --format='%H %ct'lists every commit any branch, tag or other ref leads
to, with its committer time in seconds since 1970, and the review keeps those on or after the start by comparing
whole numbers in its own code. Thengit log --no-walk=unsorted --stdinlists the workflow files that exactly
those commits added: no walk, so no side branch is passed by, and only those commits are compared with their
parents (a treeless partial clone with nothing in the window needs no tree, and reads SAFE as before).
A committer time git prints that isn't a plain whole number (git prints none for a date it can't read) makes the
repository UNKNOWN, never a commit left out.--since-as-filterwould have kept the date in git, but needs git
2.37, and Ubuntu 22.04 ships 2.34.1: every repository there would have read UNKNOWN. - A shallow repository is UNKNOWN, wherever its history is cut: "its history is cut (a shallow repository; the
newest cut is commit …, dated 10 Oct 2026 12:00 UTC), so workflows added below the cut can't be seen: run git
fetch --unshallow, then run again". Shallow meansgit rev-parse --is-shallow-repositorysaystrue, or the
shallow file git reads (git rev-parse --git-path shallow) is there and not empty; one Wardlatch can't read is
UNKNOWN too. A cut commit's date is no sign that the window is all there, because whoever makes a commit sets its
date: an earlier cut of this fix read a history cut before the start as whole, so a--depth 1clone of a tip
dated 2025, over a workflow committed on 5 Sep 2026, read SAFE, exit code 0, with the workflow file in the clone's
own tree. A shallow file can also be written into a full repository, and git then reads nothing below the commits
it names. - The SAFE line names the instant and the date it compares: "no workflow files added in commits dated on or after
30 Aug 2026 00:00 UTC (committer date) in N local repositories". The date a review line shows is the commit's
committer date in UTC, the date of the time the review compares: a commit made at 30 Aug 2026 00:00 UTC from
UTC-7, where it was still 29 Aug, shows 2026-08-30. Up to 0.1.9 the line showed the author date, and an earlier cut
of this fix the committer date in the commit's own zone. - The WordPress check's
addedSincewas already an explicit UTC instant; a test now holds the start of every
incident's window to one. - A file added more than once was listed at one commit (0.1.0 to 0.1.9, and this fix up to 3cff5c5). The review
kept one line per workflow file, at the earliest commit in the window that added it. A file added, deleted and
added again with other contents was listed once, at the first commit: Alice adds.github/workflows/ci.ymlon
5 Sep and deletes it the next day, Mallory addsci.ymlagain on 8 Oct, and the review showed only "added …
2026-09-05 Alice". It now lists every commit in the window that added each workflow file, one REVIEW line (and one
--jsonrow, with the samefileandcommitfields) per file and commit, newest first: "added 4123d9f
2026-10-08 Mallory", then "added 3fc2679 2026-09-05 Alice". With nothing added in the window, the SAFE line is
unchanged. - Still not covered: a workflow whose own commit is dated before 30 Aug 2026 isn't listed, because whoever makes a
commit sets its date (ASSURANCE_CASE.md, section 4). And a change to a workflow file that already existed isn't
listed; the check lists added files, as the campaign added them: a commit that only changes a workflow committed
before the window reads SAFE. - What to do: if 0.1.0 to 0.1.9 said SAFE for
github-actions-2026-10, run a release with this fix again, or
paste this line at the top folder of each repository, in a POSIX shell such as bash or zsh (it doesn't run as
pasted in PowerShell or cmd). It has no date in it (given one, git stops reading a branch at the first commit
dated before it) and reads what the review reads, in one command:
TZ=UTC GIT_NO_LAZY_FETCH=1 GIT_ALLOW_PROTOCOL=none git --no-pager --work-tree=. --no-replace-objects -c core.commitGraph=false -c advice.graftFileDeprecated=true log --all --full-history --root --no-renames --diff-merges=combined --diff-filter=A --name-only --no-show-signature --date=iso-strict-local --format='%h %ct %cd %an' -- .github/workflows
It lists every workflow file that any commit of any branch, tag or other ref added, each under its commit's hash,
committer time in seconds since 1970 (%ct), committer date in UTC (for convenience only) and author, whatever
the date; a file added, deleted and added again is listed under each commit that added it, as the review lists
it. **Review every line ...
- The start was a bare date. The