Skip to content

v2026.41

Choose a tag to compare

@maksimtech maksimtech released this 03 Oct 14:33
· 44 commits to main since this release
01cc29a

Added

  • Library versions, read out of the strings: a new libraries.py and a
    libraries field 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's curl.exe is the case in one line: it reports curl 8.13.0 and
    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
    banner the library states its own version and the code is here — OpenSSL 3.5.8 25 Aug 2026, libcurl/8.22.0, expat_2.8.5
    build path the source directory it was compiled in, left behind by an assert's __FILE__ — ../openssl-3.5.8/crypto/asn1/a_int.c
    assembly reference a version this file asks another file for at run time, not one it carries
    import used, 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 as 1.5.7, idn2 as 2.3.8, and the committed
    python.exe as 3.14.7. 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 use and TLSv1.3 are fourteen of
    the sixteen name-then-number strings in curl.exe, and libssh2's handshake
    banner SSH-2.0-libssh2_1.11.1 holds both numbers at once with only the second
    belonging to the library.

    Because silence is the common answer, libraries.CAVEAT is 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's curl.exe states curl 8.22.0 in its
    own banner and imports libcurl-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": null rather than dropping the key, which is the opposite
    of what tlp does and deliberately: an absent tlp means 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 2026 is in none of the four buckets. A second call would
    have read a 25 MB DLL twice, which test_the_file_is_read_for_strings_exactly_once
    now prevents.

    54 tests, of which eight run against the corpus itself. ARCHITECTURE.md gains
    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-8964 and CVE-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 under pip/_vendor/, which the Dockerfile removes along with the rest
    of the build tooling, and the fourth was openssl, which
    patchradar debian CVE-2026-84782 reports resolved in trixie at
    3.5.7-1~deb13u3 and which apt-get upgrade picked 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-2026-82560 stays, 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-2026-82560 on 2026-10-02
    still reports perl no-dsa in trixie at 5.40.1-6+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.sh keeps 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
    to ubuntu-24.04 because CodSpeed cannot run on 26.04.
    ubuntu-latest was
    Ubuntu 24.04 — read off a live run on 2026-10-02, image ubuntu24/20260927.320 —
    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 on main at 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: simulation was
    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 on main. 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-get and no sudo
    in any workflow, Python from actions/setup-python at explicit versions, no
    container: or services: jobs — and the Docker path was checked on its own by
    dispatching docker-build-check.yml against the branch, which succeeded on image
    ubuntu26/20260927.149.

    What pinning costs: nothing bumps it for you. Dependabot updates action versions,
    not runs-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.py outside 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's analyze transcript 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.exe turns
    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: --tlp on analyze and
    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 since exeradar psirt now 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 tlp field, first it 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_label holds the two
    forms to the same answer for all five labels.

    Written for real while this was added:

    > Distribution: TLP:AMBER+STRICT (TLP 2.0 — https://www.first.org/tlp/)
    > May be shared with members of the recipient's organisation only, and no further.
    
    # python.exe
    

    An unmarked report stays unmarked: no block, and no tlp key at all — not
    null, and not TLP:CLEAR. Silence from whoever ran the scan is not permission
    to redistribute. A label the standard does not define stops the command before
    the file is opened, and TLP:WHITE is refused by name with its replacement:
    TLP:WHITE belongs to TLP 1.0 and was renamed in TLP 2.0: use TLP:CLEAR instead.

    The console output is deliberately not marked. A terminal is not a document that
    travels, and a label printed where it cannot be respected teaches a reader to
    ignore it where it can.

    exeradar/tlp.py arrives by copy from apkradar, the way law_fetcher and
    first_teams do, with its own thirty-four tests; seventeen more cover the report
    and the commands.

  • The gate reads FIRST's forecast on the CVEs it already holds. EPSS is indexed
    by CVE, and SECURITY-EXCEPTIONS.toml is the one surface in this repository that
    holds CVE ids: Docker Scout names its alerts by CVE, so every accepted finding
    already has an id, a written reason and a review date. The forecast is what those
    records lacked — "no fix in any suite" accepted until December is comfortable at
    an EPSS of 0.1% and is something else at 40%.

    The forecast changes no verdict. The gate fails on a blocking alert with no entry
    and on an entry past its date, and on nothing else: exit_code takes the
    forecasts and ignores them, so the signature says they were available and did not
    decide anything. Only ids that are CVE ids are looked up —
    SNYK-DEBIAN13-GCC14-20386241 is CVE-2026-95619 and says so in its description,
    and reading prose is guessing. A CVE FIRST does not score prints "not scored by
    FIRST" rather than 0.0%, which is a real reading at the floor of the scale; FIRST
    unreachable prints nothing and the report is the one this script produced before.

    The accepted findings are listed on a passing run, worst first, because that
    is where somebody decides whether to renew a review date and nothing else prompts
    it.

    Ported from patchradar with its nineteen tests; tools/security_exceptions.py is
    shared by copy across the five, and all four copies were byte-identical before
    this.

The signature entries below do change the package, so the next release carries
a version. The Docker entry further down does not and never did: that image is
republished by dispatching docker.yml with the current version, and a tag for
it would have claimed a package change that had not happened.

Added

  • FIRST's member directory, as a source. CVE names a flaw and NVD scores it;
    FIRST is the body that says who answers for it. exeradar/first_teams.py reads
    api.first.org/data/v1/teams — public, no credentials, "access": "public" in
    every response — and returns each team with its verified address and the PGP key
    to encrypt to. Until now that lookup happened by hand, in a browser, once per
    disclosure.

    Two rules shape it, and neither is about HTTP.

    It never chooses between two organisations. A search for "hewlett" returns
    Hewlett Packard Enterprise — split off in 2015 — before HP Inc., whose name is
    on a 250 G6. Both are full members, with different addresses and different keys:

    HP Inc. PSRT hp-security-alert@hp.com 0xF46ECE7D08F8DDD9
    HPE PSRT security-alert@hpe.com a different key

    Taking the first row would send an embargoed finding to the wrong company, so
    search returns every candidate and resolve answers only on an exact name,
    handing the candidates back otherwise. Measured by mutation: a resolver that
    takes candidates[0] fails two of the tests.

    An empty answer is never an answer about the vendor. FIRST lists its members;
    most PSIRTs are not in it. "No member team" comes back as a reason that says so
    and points at security.txt instead; a timeout raises. The same split
    collectors/ keeps between "nothing found" and "could not ask".

    A fingerprint is normalised to forty uppercase hex characters or dropped — the
    directory prints some of them in spaced groups, and a malformed fingerprint
    offered as one is worse than none, because it is what a key gets checked
    against. A row with no address is dropped: it answers nothing.

    Eighteen tests. Seventeen run on trimmed copies of the live payloads; the
    eighteenth asks the real directory, and only when EXERADAR_LIVE=1 — a copy
    cannot notice a renamed field, and pgp-fingerprint disappearing would leave
    every other test green while the letters went out without a key.

    exeradar psirt <organisation> is the caller. Like verify it is meant to
    be used from a script, so the exit code carries the answer: 0 one team
    matched exactly, 1 FIRST lists no member team by that name, 2 the
    directory was unreachable or nothing it returned matches exactly. 1 and 2 are
    different answers — the first sends a reader to the vendor's security.txt, the
    second says nobody got an answer — and 2 also covers the case this command
    exists for, where printing what came back would hand a script the wrong
    company's address.

    --key prints the public key block and nothing else, for
    exeradar psirt "HP Inc." --key | gpg --import. With a fingerprint on record
    and no key block it exits 1 and prints the fingerprint, so the key can be
    fetched elsewhere and checked against it.

    Ten more tests. Two of them passed before the command existed — Typer exits 2
    on an unknown command, so an exit code alone proves nothing — and were tightened
    until they could only pass against the real thing.

Fixed

  • Mutation testing runs again: also_copy in [tool.mutmut]. The Saturday
    run died in all five Radar on 2026-10-03, before a single mutant was tried, and
    the cause was the same one each time with a different victim — here, tests/test_exceptions_epss.py could not import security_exceptions from tools/.

    mutmut copies source_paths into mutants/ and runs the suite from there,
    adding only tests/, test/, setup.cfg, pyproject.toml and uv.lock of its
    own accord. So every test that imports from tools/ or scripts/, or reads a
    file at the repository root, found nothing — and since the stats phase runs the
    suite rather than merely collecting it, one such test killed the whole run.

    The list was verified rather than guessed. mutmut 3.8 refuses to run on Windows,
    so the mutants/ tree was rebuilt by hand from mutmut's own copy rules —
    configuration.py:184 and utils/file_utils.py:66 — and the suite run inside it
    until it passed: 725 passed, 4 skipped.

    This is not the previous day's move to ubuntu-26.04: the failures are
    Python-level, inside a copied tree, and patchradar's instance dates from
    2026-09-26. The weekly cron is only what surfaced them all at once — the first
    firing since the tests that trip it were written.

  • The image Snyk scans has a fixed tag, so code scanning keeps one
    configuration for it.
    It was built as snyk-scan:${GITHUB_SHA}, and Snyk
    Container writes its own automation id into the SARIF from the image reference it
    scanned — overriding the category: given to upload-sarif. So every commit
    minted a new code-scanning configuration that nothing could ever find again, and a
    pull request was told "configurations present on refs/heads/main were not
    found"
    and could no longer be shown which alerts it had introduced.

    Measured on 2026-10-02 in apkradar, which had reached 32 of them and whose
    pull request #16 could not be diffed. This repository shows one, because its image
    does not carry the extra target Snyk names the image in. The tag is the same in
    all five, so the fix is too: the defect is there whether or not it has surfaced.

  • Two advisories of 2026-09-30 are recorded. CVE-2026-84782 (OpenSSL, DTLS
    retransmission, already fixed in Debian at 3.5.7-1~deb13u3 and waiting only on a
    rebuild) and SNYK-DEBIAN13-GCC14-20335537 / CVE-2026-102010 (open in trixie
    with no fix in any suite, and libstdc++6 is what LIEF is built on). The other
    four Radar had these entries before their tags went out; exeradar was not
    released that day, so its gate refused the next push to main instead — which is
    the gate doing its job rather than a new defect.

  • "Not valid" no longer means "we could not check". verify_signature()
    returns a bitmask of thirteen flags and this package read it as == OK;
    everything else became verified=False, which law_checker turns into
    signature_invalid — "Embedded signature not valid", severity high, cited
    against CRA Annex I Part I(2)(f) and art. 13(1), NIS2 art. 21(2)(d) and GDPR
    art. 32(1). Five of the flags say nothing of the kind:

    flag what it actually says
    UNSUPPORTED_ALGORITHM LIEF does not implement the algorithm
    CERT_NOT_FOUND the blob does not carry the signer's certificate, so there is nothing to compare against
    MISSING_PKCS9_MESSAGE_DIGEST the authenticated attributes are not laid out the way LIEF looks for. Older Authenticode blobs are not, and Windows accepts them
    CERT_EXPIRED, CERT_FUTURE the certificate's dates do not cover the moment of the check, which says nothing about the digest — and an expired certificate already has its own finding, with its own provisions

    verified is now three-valued and None means the check did not reach a
    conclusion; the flags travel with it, so a report names the reason instead of
    printing a verdict with nothing behind it. A file in that state produces a note
    saying it is a limit of the check rather than a finding about the file, and
    exeradar verify exits 2 — "could not be established" — where it used to exit
    1, the code a script reads as "this signature is bad".

  • A signature that is present and unreadable is no longer reported as absent.
    A certificate table that lies inside the file and cannot be parsed — a damaged
    copy, a format LIEF declines, a deliberately malformed blob — fell through to
    UNSIGNED: "no embedded signature and no catalog entry", about a file carrying
    14 KB of one. On Windows the catalog was consulted first and also said no, which
    made the wrong answer look corroborated. It is UNKNOWN now, with the size and
    the offset of what is there.

  • The notes no longer state a cause for UNKNOWN that may not be the one.
    Every unknown signature printed "the Windows catalog cannot be consulted on this
    platform", including on Windows, where it had been consulted, and for a file
    that simply would not parse. UNKNOWN has four causes and the note now carries
    the one that applied.

  • "Truncated" is no longer asserted as the cause. A certificate table pointing
    past the end of the file is a measurement; a cut-off download is an explanation,
    and a malformed or forged header produces the same measurement. The detail gives
    the numbers, offers the ordinary cause, and says which of the two it is has not
    been established.

  • The report claims no trust. "signed and verified" and "valid" were doing the
    work of "from a publisher this machine trusts", and no certificate store is
    consulted anywhere in this tool: a self-signed certificate verified exactly as
    well as a commercial one. What is measured is that the file still matches what
    somebody signed, and that is what the sentence now says.

    Twenty-three tests, including one that walks every flag LIEF exposes and fails
    if a future version adds one nobody has classified — an unsorted flag would land
    in "could not tell" and quietly stop being reported.

  • The runtime image no longer carries the build tooling. Docker Scout reported
    CVE-2025-47273, CVE-2026-57585 and GHSA-6v7p-g79w-8964 against
    pip/_vendor/bom.cdx.json — copies vendored inside pip and setuptools, at a
    path no dependency of this project can influence. Pinning cannot reach them: an
    explicit install of a patched msgpack adds a second copy beside the first and
    leaves the one the scanner reads exactly where it was. patchradar spent a month
    with pins in place and the same three findings open before that became clear.

    A runtime image needs none of the three: the entrypoint is exeradar, and
    neither lief, asn1crypto, typer, rich nor httpx imports pkg_resources —
    checked rather than assumed. Verified by building the image in
    docker-build-check.yml, which publishes nothing: pip, setuptools and wheel all
    absent, no pip/_vendor/bom.cdx.json anywhere, 87 packages, and the smoke test
    and CLI still passing — which is the part a removal like this could have broken.

    The three alerts close when the next image reaches Docker Hub, because that is
    what Scout reads, not the image CI builds.

  • The SonarCloud workflow ran the benchmarks and died on them. It had been
    failing since at least 2026-09-26, and not on the quality gate: pytest tests/
    includes tests/benchmarks, whose nineteen tests need the benchmark fixture
    that only codspeed.yml installs. All nineteen errored and the workflow ended
    before the scan, with 571 tests passing in the same run. Benchmarks are excluded
    here now, as in three of the other four Radar — a timing measured under coverage
    instrumentation means nothing anyway.

Added

  • A CI gate that refuses. Every other security workflow reports: snyk.yml
    carries continue-on-error, CodeQL and Docker Scout upload SARIF, and
    SonarCloud decides its quality gate after the job has already succeeded. On
    2026-09-29 all of them were green while nine high-severity alerts were open.
    security-posture.yml reads what they published and fails when a blocking
    finding has nobody's name against it; SECURITY-EXCEPTIONS.toml records the
    accepted ones with a reason and a review date. sonarcloud.yml now waits for
    its own quality gate, without which a red gate is a green job.

  • tests/docker/inspect.sh prints what the image contains instead of leaving
    it to be remembered. It settles the two claims the exceptions record rests on:
    that the build tooling is gone, and that only perl-base is installed —
    Essential, which dpkg itself depends on, so CVE-2026-82560 is ours to record and
    not ours to close. perl, perl-modules and the two perlapi entries come
    back from dpkg-query with no version at all: virtual packages perl-base
    provides, not installs. That measurement had been borrowed from patchradar's
    image and is now taken in this one.

What's Changed

  • ci: pin the runners to ubuntu-26.04, benchmarks to 24.04 by @maksimtech in #2

New Contributors

Full Changelog: v2026.40.1...v2026.41