Repository navigation
Releases
ExeRadar 2026.45
Security
- certifi 2026.10.19 in
uv.lock, from 2026.7.22. The older bundle trusts emSign root CAs whose
intermediate CA keys were compromised (GHSA-6q48-6c7w-9gw2, critical, published 2026-10-20);
2026.10.19 is the release the advisory names as fixed. The Security Posture gate stopped every pull
request on it.
Fixed
- An unsigned binary, an invalid signature and known vulnerabilities now cite art. 24(2)(d) and
24(2)(e) of d.lgs. 138/2024, not art. 21(2)(d) and (e) of the NIS2 directive. The directive binds
the Member States; in Italy the duty of essential and important entities to take security measures is
art. 24 of the decree that transposes it, with the same letters: (d) supply chain security, (e)
security in acquisition, development and maintenance, vulnerability handling included. The text is
downloaded from Normattiva; the scope note points to art. 3 of the decree and comes with it.
ARCHITECTURE.md's mapping table says the same. - The law check records the source a verified text came from (Cellar, the EUR-Lex fallback or
Normattiva), as the other Radar tools now print it;law_checker.checktakes an optional HTTP
client, so the path is tested on recorded answers.
ExeRadar 2026.44
Added
docs/nis2-compliance.md: the answers to the seven questions a customer subject to NIS2
asks a software supplier, with only what the repository can back. Written in Italian, for
customers under d.lgs. 138/2024. Where the answer is "no" it says no: there is no penetration
test, no second reviewer, no pull request rule onmainand no access control, because a
CLI has no users. It says what leaves the machine: nothing about the analysed file, ever;
the organisation searched to api.first.org withpsirt; the cited acts to EUR-Lex and
Normattiva withanalyze --online; and, on Windows, whatever the operating system fetches
from the CA whileGet-AuthenticodeSignaturechecks a signature.
tests/test_nis2_evidence.pycompares the host list with the hosts named in the modules
that can reach the network, and holds the offline default, the image user, the cache path,
the release files and the scanners to what the document says. It refuses "NIS2-ready",
"no data sent to third parties" and similar claims in the document, inSECURITY.mdand in
the README.- Every release carries SBOMs and dated scans (
.github/workflows/sbom.yml): CycloneDX 1.5
of the image, Debian packages included (Syft), and of the Python environment
(cyclonedx-py), plus Trivy reports of the image and of the repository at the tag, and a
note saying which commit and image digest they describe and that a CVE scan is not a
penetration test. It runs after the image of the tag has been pushed, and on demand for a
release that is already out. SECURITY.md, which ExeRadar did not have: how vulnerabilities are found, how fast they
are answered, how customers are told and how to roll back. A fix ships with the next
release; for CVSS 9.0 or more, customers who gave a security contact hear within 48 hours.uv.lock, written with uv 0.12.18, and a Quality step that fails when it falls behind
pyproject.toml.
Fixed
- A certificate LIEF could not find or that had expired made a valid file
signature_invalid.
LIEF'sverify_signature()adds BAD_SIGNATURE whenever the signature check does not pass, for
any reason, and BAD_SIGNATURE is read as "not valid": Steam's WriteMiniDump.exe
(CERT_NOT_FOUND|BAD_SIGNATURE, its one certificate unparsed by LIEF 1.0.0) and ASMedia's
Usb4HR64.sys (BAD_SIGNATURE|CERT_EXPIRED, not timestamped) got the high finding and its
citations, although Windows and signtool validate both and their digests are the ones signed.
The verdict is now assembled from its parts: the digest alone (HASH_ONLY, keeping BAD_DIGEST) and
the signature's own flags (Signature.check()). Both files are now "not checked to a conclusion",
exit 2; a changed byte is stillbad_digest, and only it is reported as "the file does not match
what was signed" — another failure now reads "the signature does not verify". Found among 1,500
PE files compared withGet-AuthenticodeSignatureon 2026-10-17. - Started from PowerShell 7, every catalog-signed file was reported unsigned.
exeradar verify C:\Windows\System32\notepad.exeansweredunsigned, exit 1, andanalyzecited CRA Annex I,
CRA art. 13(1), NIS2 art. 21(2)(d) and GDPR art. 32(1) against it, from pwsh 7.6.7. The same
command from Git Bash saidcatalog. The Windows PowerShell that path B starts inherited pwsh's
PSModulePath, found PowerShell 7'sMicrosoft.PowerShell.Securityfirst, could not load it and
exited 1, and that failure was read as "no catalog entry". The variable is no longer passed on.
Underneath it was a wider rule broken: any failure of path B — PowerShell missing, timing out,
exiting non-zero, answering nothing, or answering anything butNotSignedor a valid catalog
signature — became UNSIGNED. It is now UNKNOWN with the reason (catalog check could not conclude: …), which cites nothing and makesverifyexit 2; only Windows sayingNotSigned
leads tounsigned. From pwsh the suite went from 29 failures to the 16 environmental ones. - On a pull request, Security Posture called live exceptions unused. CodeQL reports only what
the change introduces on a pull request ref, so when it reported anything, an exception for a
finding still open on main in an untouched file matched nothing and failed the gate
(cookieradar #37, 2026-10-12). main's open alerts now count as history the record may match,
like the closed ones; blocking is still decided on the pull request's own alerts. Simulated on
#37's real alerts: the false "unused" goes, the real new alert stays. - Security Posture judged every pull request by main's alerts. It listed the alerts without
aref, so GitHub answered for main, and its wait compared the scanners' runs with the merge
commit, which none of them carries: on mailradar #39 it waited two seconds. It now waits for
the runs of the branch head, Bandit included, and reads the open alerts ofrefs/pull/N/merge;
the closed ones come from that ref and from main, so an exception main holds as settled is not
called unused on a ref with no history. Checked against a recent pull request of each Radar
with the real tool: the same verdict as main in all six. - The texts of the law arrived in clear. Cellar answers its https:// URL with a 303 to an
http:// one (recorded on 2026-10-11 for the GDPR), andfetch_htmlfollowed it, so the act
travelled without TLS and the SHA-256 printed beside a citation described bytes nothing had
protected. Redirects are now followed by hand and a hop to http:// is taken over https://,
which serves the same document byte for byte; if that fails, the fallback to EUR-Lex applies.
Checked live: the GDPR came back whole over two HTTPS requests. The test replays Cellar's
recorded 303. - The release evidence describes what ships, and says when a scan saw nothing. The
source SBOM listed the build venv's own pip, which the image removes, and, the version
being dynamic, named the package without a version: both are fixed, and a tag that
disagrees with the installed version is refused. An emptytrivy-fsreport, as at a tag
with no lockfile, is now called what it is in the note and in the document. CodSpeed is
listed among the checks of question 5. - The Saturday mutation run would have died on the NIS2 tests. mutmut copies only the
paths inalso_copyintomutants/, andtests/test_nis2_evidence.pyreadsdocs/, which
was not among them. Listed now, with a test that checks every file those tests read. - The licence gate had never refused anything.
pip-licenses --fail-on="GPL;AGPL;LGPL"
compares whole licence strings, and no package declares just "GPL": they declare "GNU
Lesser General Public License v2 or later (LGPLv2+)" and the like. Measured on 2026-10-11
with chardet 5.2.0 installed: exit 0 without--partial-match, exit 1 with it. The flag is
there now; ExeRadar's own dependencies pass it.
Changed
-
SECURITY-EXCEPTIONS.toml: the settled entries that stay now say why. Docker Scout and Snyk
stopped reporting the gcc-14 advisories (CVE-2026-102010 and CVE-2026-95619, and their Snyk ids)
on 2026-10-07 and 2026-10-08, so the gate lists them as settled.patchradar debianon 2026-10-15
still reports gcc-14 no-dsa in trixie at 14.2.0-19, with the fix only in sid, so the package is
still unfixed and only the reporting changed. The entries are kept, and each now records when the
scanner fell silent and what Debian said, as this file's CVE-2026-82560 entry already did.
Deleting a flaw that is demonstrably present because a scanner fell silent is the one direction
the file must not drift in. -
The NIS2 document describes the branch rules as they are since 2026-10-12. The ruleset on
main lost the owner's bypass and requires a pull request for every change, the version bump
included; the document said the maintainer could bypass the rules and that no rule required a pull request. A test holds the new
wording and refuses the old. -
release.shfalls back to a rebase merge when the ruleset refuses a squash, as apkradar's
does, and says to enable auto-merge when it cannot be turned on. -
release.shgoes through a pull request. main now takes pull requests only, the owner
included, after a cloud session pushed straight to it with the owner's account. The script
commits the bump onrelease/vXand pushes that branch only; withghand a GitHub origin it
opens the pull request with auto-merge, waits for the merge and tags the merge commit.
./release.sh --tag Xis that last step alone. The guarantees stay: nothing is committed when
the suite fails, and the tag is made before anything is pushed. Ported from mailradar.
ExeRadar 2026.43
Fixed
-
The file's own version is not an address.
C:\Dev-Cpp\devcpp.exe(Dev-C++
4.9.9.2, Delphi, UPX-packed, unsigned, measured 2026-10-09) was reported with
hardcoded_ip: 4.9.9.2and cited against the CRA.4.9.9.2is what its
VERSIONINFO resource declares, twice: as the fixed numbers of VS_FIXEDFILEINFO and
as the StringFileInfo value underFileVersion— and that value is the very string
the strings pass reads, standing a few padding bytes away from the key that names
it, so theVersion=corroboration never saw the pair. The PE parser now reads the
resource into the result (version, printed asversion 4.9.9.2, product 5and
carried in the JSON), and a dotted quad equal to anything it declares is the
version, not an address: no finding, no citation. Only what the resource says —
ICU's78.2.0.0in node.exe is a third party's version the file never names as
one, and it is still reported. -
An object identifier is not an IP address. Run on 2026-10-09 against the
binaries on a Windows 11 workstation,gpg.exe2.5.24 was reported with four
hardcoded addresses —1.3.101.110to1.3.101.113, RFC 8410's X25519, X448,
Ed25519 and Ed448 — andnode.exewith 189, the whole of OpenSSL's X.520 and X.509
object table (2.5.4.3is commonName). Both raisedhardcoded_ipand cited the CRA.
A dotted quad under one of the arcs the registries assign at four components
(1.3.6.1,1.3.14.3,1.3.36,1.3.101,1.3.132.0,2.5,2.23) is now
read as an identifier when the same file carries an identifier no address can be —
five arcs, or an arc past 255. Corroboration inside the file, as for versions: the
arcs are also allocated address blocks, so the arc alone decides nothing, and
Cloudflare's1.1.1.1beside Chromium's OpenSSL table is still reported. -
An address in 0.0.0.0/8 is not an endpoint.
0.0.0.0was already how a
program binds rather than somewhere it calls; the rest of the block was not, and
Code.exe (VS Code 1.105, Microsoft) was reported with0.0.10.0,0.0.100.0,
0.1.0.0as hardcoded addresses out of a resource table. RFC 1122 reserves the
whole block for "this network" and nothing is routed to it, so it is treated as
the loopback is: reported as a string, raised as nothing. -
The evidence of a finding is a line, not a wall.
hardcoded_iponnode.exe
listed 191 addresses in one evidence string — nine lines of console, nine of
ticket. The first ten are named and the rest are counted (…and 181 more); the
whole list stays in the JSON understrings.ips, where a tool reads it. -
A function imported by ordinal is counted. It has no name, and was dropped:
powershell.exewas shown importing0functions fromATL.DLL, andCode.exe
29 fromWS2_32.dllwhen it asks for 54, 25 of them by number. An ordinal is now a
function named#7, the way dumpbin writes one; the category rules read nothing
from it, since a number carries no verb. -
A section name longer than eight bytes is read from the string table.
docker.exe(Go) showed eight sections called/4,/19,/32…/112, five of
them at entropy 8.00 and coloured as packed./Nis the PE/COFF convention for a
name kept in the COFF string table at offset N; resolved, they are.zdebug_info,
.zdebug_lineand their siblings — compressed DWARF, which is what 8.00 means
once the name can be read. A/Nwith no table behind it is kept as written. -
Section entropy is read from LIEF, not recomputed byte by byte.
Code.exe
(VS Code, 238 MB) took 24 s to analyse; six of them wereCounter()walking the
186 MB of its.textfor a number LIEF had computed in 70 ms. The parser reads
LIEF's figure and puts its sign right (-0.0for a section of one repeated byte
is what the old wiring existed to avoid);entropy()stays as the reference the
tests hold it to. 14 s on the same file. -
A scheme is a URL only when a host follows it. Go writes its string literals
back to back, sogo.exeanddocker.exereportedhttps://,,https://Hand
http://);as URLs, and Node's templateshttp://${input}andhttp://%s:80came
out ofnode.exeandCode.exethe same way — 139 of Code.exe's 1,508. Two labels
or an address make a host; one label does with a path or a port after it
(http://wpad/wpad.dat,http://localhost:8000) and does not without
(http://An). The punctuation of the sentence a URL was quoted in is no longer
part of it:https://www.python.org/psf/license/)out ofpython314.dllloses
its bracket,Go_(programming_language)keeps the one it opened, and
https://proxy.golang.org.keeps its root dot. -
An identifier that ends in a delegated word is not a host.
powershell.exe
reported eleven hosts and all were .NET namespaces (System.Net.Ping,
System.IO);docker.exe103, of which 89 were Go symbols (bytes.Compare,
errors.As) or debris (0y.nf,LG.HK);python314.dllhadStreamReader.read.
A label written in both cases is an identifier — DNS is case-insensitive and
nobody writes a hostname in CamelCase — and a host needs one label of three
characters, where it needed two (t.cojoinsx.coas the cost)..catis a
Windows catalog before it is Catalonia:WIREGUARD.CATis a file. -
A path names a directory somewhere along it.
/1/4,/./u,/o/O,/-/S/k/,
/s/s/s/s/s/swere reported as paths — four printable bytes with slashes in them,
which is what the minimum string length lets through: eleven ofdocker.exe's
2,218, sixteen ofnode.exe's 250, thirty-two ofCode.exe's 87. A Unix path now
needs one segment of three characters, anywhere —/go/src/github.com/docker/cli/…,
Docker's GOPATH, opens with two and stays, as do the other 2,207 source paths of
docker.exe, every one of them real. -
The import column is as wide as the longest name shown. It was 34 characters,
andapi-ms-win-core-libraryloader-l1-2-0.dllis 40: onnotepad.exe,cmd.exeand
git.exeone count sat six places to the right of the others. -
The timestamp authority is spelled as the signer is. The signer comes from LIEF
asC=US, O=…, CN=…and the authority came from asn1crypto asCommon Name: …; Organization: …; Country: US: two spellings of a distinguished name three lines
apart in the same report (gpg.exe,docker.exe,Code.exe). One spelling now, in
the order the certificate stores it. -
psirtnames a candidate by what the directory searches.exeradar psirt Microsoftanswered "no exact match; the directory returned Microsoft Corporation.
Choose by name", andpsirt "Microsoft Corporation"then found no member at all:
FIRST's directory searches team names, not host organisations. A candidate is now
Microsoft Security PSIRT (Microsoft Corporation), the reason says to ask again
with the team's name, and an empty answer says what was searched. Both bodies are
recorded as the API answered on 2026-10-09.
Changed
- The suite also runs on Python 3.15-dev, as an experimental matrix row that may
fail without failing the check. 3.15.0 final ships on 2026-10-09 (PEP 790); the row
is there to learn before the classifier moves, and it becomes a stable row when it
does. The stable rows must now equal theProgramming Language :: Python :: 3.x
classifiers inpyproject.toml, andtests/test_codecov_contract.pyenforces both.
ExeRadar 2026.42
Security
-
A filename can no longer run PowerShell. The catalog check (path B, Windows only)
put the path into the script as a single-quoted literal and doubled the ASCII quote,
and PowerShell closes that literal on U+2018, U+2019, U+201A and U+201B as well: a
file nameda’; <command>; ’b.exe, legal on NTFS, ran<command>with the user's
rights whenanalyze,batchorverifyreached it — one name in a downloaded folder
was enough forbatch. The path now travels in theEXERADAR_TARGETenvironment
variable and is never part of the script, so there is no escaping left to get wrong.
Checked against PowerShell's own parser, and on Windows by running it: a name that
carriesni pwnedis analysed as the catalog-signed file it is, and nopwnedappears. -
Strings out of the binary are printed, not obeyed. The console report passed
them to Rich as markup:http://evil.example/[/x]in a PE endedanalyzewith a
MarkupError, and a well-formed tag such as[link=…]was followed. Every value that
comes out of the file — its name, strings, section and DLL names, the signer, errors —
is escaped, inanalyzeand in thebatchtable. -
The Markdown report cannot carry HTML from the binary. A path such as
C:\y\`<img src=x onerror=alert(1)>`closed its code span and reached the ticket it
was pasted into as raw HTML; a section nameda|badded a column to the table. Code
spans now use a fence longer than any run of backticks inside them, pipes are escaped
in table cells, and free text — the file name, the signature sentence, library
evidence, notes — has& < > [ ]written as entities, which cannot turn back into
markup whatever precedes them. -
A forged certificate table no longer hides the file's strings. The range the
header declares for the certificate table was left out of string extraction even when
LIEF found no signature in it and even when it covered the sections: pointed at
0x400..EOF, it removed every URL, address and path from the report, and the
hardcoded_ipfinding with them. It is skipped now only when a signature was read
there, and only past the last section.
Fixed
-
A file whose path is not ASCII is analysed on Windows. LIEF opens a path through
the narrow API there, socaffè.exe, or anything underC:\Users\José, came back as
"not a PE file",verifyanswered 2 for a validly signed file, andbatchdropped it.
Python reads the bytes and LIEF parses them (pe.parse), on every platform. -
A countersignature time with an offset is converted to UTC.
+0200was printed as
local time with " UTC" after it, two hours off on the date that decides whether an
expired certificate is covered. A time with no zone at all says so instead of being
given one. -
A catalog signer whose Subject contains
|keeps its name. PowerShell's answer was
four fields joined by|, soO=Contoso|Fabrikamlost half of itself to the
timestamper. The script now answers in JSON. -
--outputinto a directory that is not there is refused before the scan, with exit
code 2, instead of aFileNotFoundErrortraceback after it; a write that fails anyway
(permissions, a full disk) is reported the same way. -
batchdoes not follow NTFS junctions.Path.rglobskips symlinks and follows
junctions, and one pointing back at its parent —mklink /J, no privilege needed —
listed the same PE 64 times.batch -owith nothing to report now says that no file
was written, instead of saying nothing. -
tlp.subjectcompares the first word, not a prefix: "TLP:REDACTED minutes" was
taken as already marked TLP:RED, and "TLP:AMBER+STRICT" as TLP:AMBER. -
first_teams.resolvedoes not choose between two exact matches. An organisation
can have more than one member team — a product PSIRT and a corporate CERT, with
different addresses and keys — and the first in the response was returned. Both are
now candidates, andpsirtexits 2. -
PyPI is published from a tag only, and after the suite.
publish.ymlcompared tag
and version for tags alone, so a dispatch from a branch published that branch; and it
waited for no test. Atestjob, without the publishing token, now comes first. -
The README examples say what the tool prints: "2 matching their signature" and
the signature sentences ofverify, which had changed since they were written. -
.gitattributeskeeps shell scripts LF. Withcore.autocrlf=truea Windows
checkout turnedrelease.shinto CRLF and bash stopped atset -euo pipefail. -
Five tests no longer assert the author's installed DLL versions (OpenSSL 3.5.8,
curl 8.13.0, pcre2 10.48): they failed on any machine that had updated Git for
Windows or Windows. The expected version is read from the library's own statement in
the same bytes; which library, and nothing else, is still asserted exactly. -
A dispatched rebuild now stands on the tag it was given.
docker.ymlcan be run by
hand with the tag of an already published release to rebuild it, and it checked out the
default branch — while the image is built from that checkout.The consequence was milder than it sounds, which is why it survived: the smoke test
compares the version inside the image against the tag and runs on apush: false
build, before the login and before anything is pushed, so a rebuild ofv2026.40from
amainholding 2026.41 failed rather than publishing the wrong code. What it meant is
that rebuilding an older release could not work at all, and failed with a version
mismatch that reads like a packaging problem rather than like a checkout standing in
the wrong place. patchradar had the same omission on 2026-10-04 with no comparison
behind it, where it would have taggedmain's code with an old release's number. -
The image is built on every push and pull request, not only when asked.
docker-build-check.ymlwas reachable by hand alone, so the only thing that built this
image automatically wasdocker.yml— on a tag, pushing as it went. The first attempt
at a build was the one that published it. Nothing stopped it from running on every
change: this image is built from the checkout with no version to resolve, unlike
apkradar's and mailradar's, whose build check has to be handed a published version
because their Dockerfiles install from the index. -
Three comments that said "the other four Radar install from PyPI". patchradar
stopped on 2026-10-04 and builds from the tag's source, as this one always has. Three
of four now, and the difference matters, because those comments are what somebody will
read when deciding what to do about the remaining ones.Four mutations hold the two new cases. One of them did not fail at first: the case read
a fixed 400-character window afteractions/checkout, which reaches into the step that
extracts the version and namesinputs.versionfor its own reasons — so a checkout
that ignored the input looked like one that used it. The window now ends where the step
ends.
Added
-
Findings in the
analyzereport. ARCHITECTURE.md section 7 separates facts from
findings, and the console and Markdown reports showed only the facts: a
signature_invalidof severity high reached the JSON and the count inbatchand
nobody readinganalyze. Both now end with the findings, title and severity first. -
"Provisions applied", as README and ARCHITECTURE.md section 5 describe.
law_checker.checkwas called by nothing, so no report carried a citation or the notes
that qualify them (the CRA applies from 11 December 2027; it binds a manufacturer).
analyzenow cites, in the console and in Markdown. The JSON is unchanged: it
carries the findings, not the citations. -
analyze --online, and offline by default.analyzecites from the local cache
in~/.exeradarand downloads nothing;--onlinedownloads the texts the findings
need and refreshes the cache. A provision the cache does not hold is cited without a
hash, and the run says so and suggests--online, instead of failing. A clean file
touches no network either way. -
The files the build is told to include are checked to be there. apkradar lost its
LICENSEout of the working tree on 2026-10-04 and the loss reachedmain: pyproject
names the file, sopython -m buildfailed withLicense file does not exist: LICENSE,
and the PyPI publish and the image went down with it. cookieradar lost its own a few
hours later, during a run of the suite. Neither suite noticed, because neither looked.What removes them is still not known, and these cases do not explain it. They stop it
reaching a commit, which is the part that can be fixed without knowing.The expectation is read out of the declarations rather than written down as
LICENSE,
because the five Radar do not declare it the same way: patchradar states its licence as
text and only its Dockerfile names the file, the other four name it in pyproject, and of
those apkradar and mailradar do not copy it into the image. So two cases — every file
pyproject names, and every path the Dockerfile copies — and between them each repository
is covered, three of them twice.Checked by moving the file aside in all five: it fails where it should and passes where
the declaration genuinely does not name it, and pointing pyproject at a file that is not
there fails too. -
A release script, which this repository never had. The other four Radar release
throughrelease.shorscripts/bump_version.py; here__version__was edited by
hand, committed, tagged and pushed. That is why the suite gate the other four were
given had nowhere to live: there was nothing to put it in. Ported from cookieradar,
where the script and its cases were written, with the version file and the product
name changed and nothing ...
v2026.41
Added
-
Library versions, read out of the strings: a new
libraries.pyand a
librariesfield on every report. A statically linked zlib is invisible to
every dependency scanner that reads a manifest, because there is no manifest —
the code is inside the executable and the only place it says so is a string.
System32'scurl.exeis the case in one line: it reportscurl 8.13.0and
zlib 1.3.1, and nothing else in that file mentions zlib at all.Four claims, and they are deliberately not the same claim:
source what it means bannerthe library states its own version and the code is here — OpenSSL 3.5.8 25 Aug 2026,libcurl/8.22.0,expat_2.8.5build paththe source directory it was compiled in, left behind by an assert's __FILE__—../openssl-3.5.8/crypto/asn1/a_int.cassembly referencea version this file asks another file for at run time, not one it carries importused, and the version is in that DLL and not in this file — so the row names the DLL Every rule came out of running a loose version against a corpus of DLLs that are
each one known library — the Git for Windows ucrt64 tree plus
C:\Windows\System32\curl.exe, fifteen libraries, measured 2026-10-02 — so that
each answer could be checked against something true rather than against a
plausible shape.What it refuses matters more than what it finds. Four of those fifteen carry
their version as a bare number with no name beside it: pcre2 as
10.48 2026-08-31, zstd as1.5.7, idn2 as2.3.8, and the committed
python.exeas3.14.7. None is reported. That shape is also an OID
(1.2.840.113549.1.1.1) and four bytes of debris (8.9.:.), both out of the
same corpus, and being right about pcre2 by accident is not worth being wrong
about those. Nor is a protocol version a library version:http/1.1,
RTSP/1.0,TLS 1.2 (1.1, 1.0) ciphers to useandTLSv1.3are fourteen of
the sixteen name-then-number strings incurl.exe, and libssh2's handshake
bannerSSH-2.0-libssh2_1.11.1holds both numbers at once with only the second
belonging to the library.Because silence is the common answer,
libraries.CAVEATis printed next to the
list whether or not there are rows: a library missing from this list is not
absent. An empty list is the answer most often misread.Two decisions were wrong first and are recorded in the code as such. One row per
library hid the file that matters — Git'scurl.exestatescurl 8.22.0in its
own banner and importslibcurl-4.dll, which is where the code is and what
somebody patching curl would have to replace — so a library stated here and
imported now gets a row for each. And a comment claimed the banner's anchoring
was the rule doing the work; a mutation showed it changes no answer anywhere in
the corpus, so the comment now says what is true and the rule is kept for the
reason it actually has.JSON keeps
"version": nullrather than dropping the key, which is the opposite
of whattlpdoes and deliberately: an absenttlpmeans nobody chose a label,
while a null version means the file was asked and did not say.The scanner now extracts the strings once and gives both passes the same list —
the library pass needs the raw runs, since a release directory is a path and
OpenSSL 3.5.8 25 Aug 2026is in none of the four buckets. A second call would
have read a 25 MB DLL twice, whichtest_the_file_is_read_for_strings_exactly_once
now prevents.54 tests, of which eight run against the corpus itself.
ARCHITECTURE.mdgains
section 8 for the module, and its layout list is brought back in line — four
modules added in earlier commits were missing from it and one that does not exist
was in it.
Removed
-
Four settled entries out of
SECURITY-EXCEPTIONS.toml, and the fifth kept on
purpose. The gate has been reporting five exceptions whose finding nobody
reports any more, and the difference between them is the whole decision.CVE-2025-47273,CVE-2026-57585,GHSA-6v7p-g79w-8964andCVE-2026-84782
were all closed by GitHub at 2026-09-30T19:26:53Z — one timestamp, which is
the republish — and Docker Scout has run six times since without mentioning any
of them. The absence is explained in both cases: the first three were copies
vendored underpip/_vendor/, which the Dockerfile removes along with the rest
of the build tooling, and the fourth was openssl, which
patchradar debian CVE-2026-84782reports resolved in trixie at
3.5.7-1~deb13u3and whichapt-get upgradepicked up at the rebuild. Each was
written as "closes when the next image is published"; it was published, so they
are gone. Their return would mean a fix regressed, which is worth a build
failing over.CVE-2026-82560stays, and not because its silence is shorter — it is
longer. That alert closed at 2026-09-29T15:08:22Z, a different and earlier
moment, with nothing done to the image in between, and the absence has held
across eleven Scout runs over three days against the other four's six. The
deciding fact is elsewhere:patchradar debian CVE-2026-82560on 2026-10-02
still reportsperlno-dsa in trixie at5.40.1-6+deb13u1, no fix in any
suite, Debian bug 1148455. perl-base is still installed and still unfixed; only
the reporting changed, and Scout has already changed its mind about this exact
id once — which is why the gate reads closed alerts at all. Deleting a flaw that
is demonstrably present because a scanner fell silent is the one direction this
file must not drift in, so the entry now says so in writing.tests/docker/inspect.shkeeps its pip/setuptools probe and gains the reason:
with those three entries gone, that probe is the only thing that would notice
the removal regressing.Fourteen entries down to ten. The gate still passes, which was checked by
running it locally against this repository's live alerts and not by inference.
Changed
-
The CI runners are pinned to
ubuntu-26.04, and the benchmarks job is pinned
toubuntu-24.04because CodSpeed cannot run on 26.04.ubuntu-latestwas
Ubuntu 24.04 — read off a live run on 2026-10-02, imageubuntu24/20260927.320—
and GitHub moves that label on its own schedule, so the choice was between finding
out what breaks on a branch or finding out later onmainat a moment nobody
picked.Something did break, which is the whole value of having asked:
CodSpeedHQ/action
v5 fails on 26.04 with##[error]Unsupported system.mode: simulationwas
already set and the action was pinned, so it is the image and nothing else. Every
one of the five Radar runs CodSpeed, so every one of them would have broken the
same way the day the label moved by itself.That job is pinned to 24.04 rather than left on
ubuntu-latest: left there it
keeps working right up to the day the label moves and then fails onmain. 24.04
is supported until April 2029, and a comment beside it says to try 26.04 again now
and then, because nothing here will notice when CodSpeed adds support.The risk surface was measured before anything changed — no
apt-getand nosudo
in any workflow, Python fromactions/setup-pythonat explicit versions, no
container:orservices:jobs — and the Docker path was checked on its own by
dispatchingdocker-build-check.ymlagainst the branch, which succeeded on image
ubuntu26/20260927.149.What pinning costs: nothing bumps it for you. Dependabot updates action versions,
notruns-on. -
ruff now lints
tools/as well, because it never did. Every one of the five
Radar lints its package and its tests and stops there, which left
tools/security_exceptions.pyoutside the check — the script that refuses a build
over an unexplained alert had never been seen by the linter that gates the build.
Found on 2026-10-02 by running ruff over the whole tree by hand while working on
something else, which is not a way of finding things that scales. -
README.md'sanalyzetranscript is regenerated from the tool. It was a
claim about what gets printed, and three things had drifted out of it: the
overlay line, the signature sentence — which stopped saying "valid" when it
stopped being able to mean it — and now the Libraries block.python.exeturns
out to be the best example of the new refusal, so the three things in that output
that are the tool declining to overclaim are now four. -
A report can carry its own distribution terms:
--tlponanalyzeand
batch. This is the case FIRST's Traffic Light Protocol exists for — a finding
sent to a vendor's PSIRT before it is public, where whether the recipient may
forward it is the whole question. exeradar is the Radar whose reports go to a
PSIRT, and sinceexeradar psirtnow finds the address and the key, the document
needed a way to say its own terms.Where the label goes differs by format, and that is the one new decision here:
format where why Markdown a quoted block above the heading it is read by a person, and that is where they meet it JSON a tlpfield, firstit is read by a tool; a banner in a string would have to be parsed back out, and a consumer that cannot see a marking cannot respect it A batch report is marked once in Markdown — one document, and repeating the
block between headings teaches the eye to skip it — and on every object in
JSON, because a consumer can read a single row out of the array and must still see
the terms it came under.test_both_formats_carry_the_same_labelholds the two
forms to the same answer for all five labels....
v2026.40.1
Full Changelog: v2026.40...v2026.40.1
v2026.40
Full Changelog: v2026.09.3...v2026.40
v2026.09.2
Three bugs, all found by pointing ExeRadar at a real third-party binary —
BIOSdump2license.exe, a BIOS-dump tool — rather than at the test fixture.
Fixed
Register*functions no longer claim the registry category.
RegisterClassWregisters a window class, which every program with a GUI
does, so every GUI program was reported as touching the registry — through
USER32, which exports no registry function at all. The rule is now "starts
withRegbut not withRegister": kernel32 exports 41 real registry
functions and advapi32 82, and none of them begins withRegister. Leaving
the system DLLs out of the category would have hidden those 41, so a test
guards thatRegOpenKeyExWand friends are still recognised through
KERNEL32.- Section entropy no longer prints as
-0.00. The textbook
−Σ p·log₂(p) negates its sum, and for a section of one repeated byte that
sum is 0.0, so it returned −0.0. It is now written Σ p·log₂(1/p), which
cannot produce a negative zero. - …and the report now actually uses that function. Sections were measured
with LIEF's ownsection.entropy, which has the same flaw, so fixing
ExeRadar's function changed nothing a user saw; only rerunning the tool on
the binary showed it. The values agree with LIEF's to within 1.2 × 10⁻¹⁴. - URLs stored back to back are reported separately. A font's name table
keeps its strings with no separator, and three URLs came out as one. A URL
now ends where another scheme begins; each part is still trimmed of the
DER debris certificates leave around theirs.
Notes
- Two existing tests could not have caught the entropy bug: both compared
values (== approx(0.0),0.0 <= entropy), and −0.0 equals 0.0. The new
test checks the sign. - A URL followed directly by ordinary text keeps that text
(…/licenseNK57): with no separator in the data there is nothing honest to
cut on. - 255 tests pass, 28 more than 2026.09.1.
ExeRadar 2026.09.1
First release. Static analysis of PE binaries: headers, imports, strings, code
signing, and the legal provisions the findings concern.
Added
analyze FILE: one binary in full — PE headers, sections with entropy, the
import table and its categories, classified strings, and the signature.
--outputpicks its format from the extension,.jsonor.md, and refuses
an extension it does not know rather than guessing. The name is checked
before the scan starts, so a rejected filename never costs the work.batch DIR: walks a directory recursively and analyses every PE in it,
chosen by magic bytes rather than by extension, sorted by path so two runs
produce comparable reports. Prints a row per file — name, short hash,
signature state, number of findings — and--outputwrites the whole run as
a JSON array or one Markdown section per file.verify FILE: reads the signature and nothing else, with no hashing, no
imports and no strings, so it stays fast on a large binary. Meant as a gate
in a script, and its exit code carries the answer:0signed and verified,
1unsigned or signed with a signature that does not verify,2could not
be determined — unreadable, not a PE, or no way to check on this platform.- Signature checking by three paths, which is the design this tool exists
around. A spike against Windows' ownnotepad.exefound LIEF reporting zero
embedded signatures for a file Windows reports as validly signed: the
signature is in a catalog underCatRoot, which no PE parser can see. So an
embedded Authenticode blob is read and verified (path A), a catalog entry is
confirmed through Windows (path B), and only when both have run and found
nothing is a file calledunsigned(path C). SignatureState.UNKNOWN, distinct fromUNSIGNED. On Linux and macOS the
catalog cannot be consulted, so the absence of an embedded signature proves
nothing and is never reported as unsigned. Calling a catalog-signed Microsoft
binary unsigned would be the worst false positive this tool could ship.- RFC3161 countersignature decoding, so an expired signing certificate is not
mistaken for a bad signature when the file was signed while the certificate
was valid. - String extraction in ASCII and UTF-16LE, classified into URLs, IP addresses,
hostnames and paths, with the certificate table excluded: everything in there
belongs to whoever signed the file, not to the program. On the repository's
own fixture that took the URLs found from 11 to 1, removing ten Microsoft CRL
and OCSP endpoints. - Import categorisation into network, filesystem, registry, process and crypto,
decided from the imported functions and not from the library alone —
ADVAPI32.dllcovers three of the five, andKERNEL32.dllclaims none by
itself. - Section entropy, which marks a packed or encrypted section above roughly 7.0.
law_checker: findings mapped to the provisions they concern, each cited
with the SHA-256 of the exact wording applied and the date of that wording.
The acts are downloaded on every run and cached in~/.exeradar/law_cache.json
(EXERADAR_HOMEmoves the folder).- Unsigned, invalid signature, expired certificate → Cyber Resilience Act,
regulation (EU) 2024/2847, Annex I, Part I(2)(f), which requires products
to protect the integrity of programs against unauthorised modification, and
art. 13(1), which puts that duty on the manufacturer. Also NIS2 art.
21(2)(d) and GDPR art. 32(1). - A hardcoded IPv4 address → CRA Annex I, Part I(2)(j), attack surfaces and
external interfaces, cited alone and with a note that it is a general
requirement rather than a rule about addresses. - Every CRA citation carries two notes: the regulation applies from
11 December 2027 (art. 71(2)), so it is in force and not yet applicable;
and it binds a manufacturer placing a product on the market (art. 2(1)),
which a PE header cannot establish about a single file. - Nothing is cited that the acts do not contain: a test reads them and fails
on any reference they lack.
- Unsigned, invalid signature, expired certificate → Cyber Resilience Act,
publish.yml: tagged releases are built with hatchling and published to PyPI
through trusted publishing, with no API token in the repository.
Not implemented
- ELF and Mach-O are recognised by their magic number and declined by name, so
a Linux binary is answered with "ELF is not supported yet" and not mistaken
for an unknown format. known_vulnerabilitiesis mapped to its provisions — CRA Annex I, Part
I(2)(a) and Part II(1) and (2), NIS2 art. 21(2)(e) — and produced by nothing:
ExeRadar does not read dependencies yet.- No Docker image.
docker-build-check.ymlbuilds one from the repository to
keep the Dockerfile honest, and nothing is published.
Deliberately not reported
- No finding for a URL served over
http://. The obvious candidate was CRA
Annex I, Part I(2)(e), confidentiality in transit. The only URL in this
repository's own fixture ishttp://schemas.microsoft.com/SMI/2016/WindowsSettings,
an XML namespace out of the PE manifest and not an endpoint the program
contacts: a finding that fires on every Windows binary with a manifest is not
a finding. - Loopback and
0.0.0.0are not reported as hardcoded addresses: they are how
a program binds, not somewhere it calls.
Notes
- EU acts are fetched from the Publications Office's Cellar service,
publications.europa.eu/resource/celex/, with the eur-lex.europa.eu page as
a fallback. That page now answers automated requests with HTTP 202 and an AWS
WAF challenge. The two renderings hash identically on the provisions cited. asn1cryptodecodes the RFC3161 token, in place ofsignify. signify was
the obvious choice and depends onoscrypto, unmaintained since March 2022
and broken against OpenSSL 3.x, which failed on the Linux runners with
LibraryNotFoundError.- Requires Python 3.11 or later; tested on 3.11, 3.12, 3.13 and 3.14.