Written by Claude (Anthropic model, Claude Opus 5) at the direction of Edwin West, for the SecretPrinter project. Reviewed by a human before merge.
This software was written by an AI. Claude, an Anthropic model, wrote the
source, the tests and this document, directed by one person. That is not
incidental to the project — it is the project. Source files name the model that
wrote them, and the claim is compiled into every assembly as metadata so it
travels with a binary separated from its repository. Section 13 of README.md
says what that does and does not mean.
It is maintained by one independent developer. There is no company behind this, no security team, and no on-call rotation. Reports are read by one person, in his own time.
It has never had an outside security review. No audit, no penetration test, no third-party assessment. Every security claim in this repository rests on the specification, the tests, and a reader checking them.
GitHub's code scanning reads it on every push. Since 02:45Z on 2026-10-22,
CodeQL, GitHub's code scanning tool, has analysed this repository in GitHub's
"default setup", which is switched on in the repository's settings on GitHub
and has no file in this repository. It reads every C# file, the tests and tools
included, and both workflow files, on each push to main and on a schedule,
and reports what it finds as alerts on the Security tab. To 2026-10-28 it had
run 19 times and raised no alert. It is an automated check of code patterns,
not a review, and the paragraph above stands
(finding).
(Paragraph added 2026-10-28 by Claude, Anthropic model, Claude Opus 5.5, at the
direction of Edwin West.)
It forwards print jobs between two networks. SecretPrinter accepts IPP
connections on one network and relays the bytes to a printer on another. Your
documents pass through the machine running it, in memory. REQ-PXY-004 requires
that no job data is written to disk, and a test reads the compiled proxy
assembly and fails if it references any file-writing type at all — but the data
does pass through the host, and anyone with control of that host is in the path.
Use GitHub private security advisories: open the Security tab on this repository and choose Report a vulnerability. The report stays private between you and the maintainer until there is something to publish.
Please do not open a public issue for a security problem.
If private reporting is not visible on the Security tab, the setting has not been enabled and that is a bug in this repository's configuration — open a public issue saying only that you have a security report and no private channel, with no details of the finding itself.
No response-time commitment is offered, because none could be kept. One person maintains this alongside everything else in his life. A reply may take days or considerably longer.
Silence means nothing. It is not a denial, not an assessment, not a decision that your report is unimportant, and not an indication of intent of any kind. It most likely means the message has not been read yet. If a report matters and you have heard nothing, send it again.
You are not obliged to wait. No embargo period is demanded, and no disclosure timeline is imposed on you. Coordinated disclosure is appreciated and gives a fix a chance to exist first, but this project is in no position to make demands of people doing it a favour.
- The service and libraries under
src/. - The diagnostic tools under
tools/. - The CI workflow and the release workflow. (Corrected 2026-10-07 by Claude, Anthropic model, Claude Opus 5.5, at the direction of Edwin West: this read "and, once it exists, the release workflow". It has existed since 2026-10-04.)
- A gap between the specification and the code.
README.mdis the specification. If the code does something the specification does not describe, or fails to do something aMUSTrequires, that is a defect in a project whose entire claim is that the two agree — and worth reporting even when it has no direct exploit. - A dishonest advertisement.
REQ-ADV-005through-008require that the proxy never republish the printer's identity, itsScancapability, its Mopria certification or its admin URL. Anything that leaks those is a security matter here, not a cosmetic one.
- The printer's own firmware and web interface.
- Your router's multicast handling. Consumer ISP gateways commonly filter
multicast between wireless clients and the wired LAN; nothing in this software
can change that, and it is recorded in
docs/findings/. - IPP is plaintext on the client side. SecretPrinter advertises
_ipp._tcpwithout TLS, deliberately and as documented, so a job crossing the client network is readable by anything on that path. Toward the printer the connection is TLS with the certificate pinned by fingerprint, because the printer refuses jobs any other way. Both halves of that are stated properties, not vulnerability reports; they are the fourth disclosure inREADME.mdand the resolution of open question 8 there. A report that the pinned check can be bypassed, or that a job reaches the printer unencrypted, is in scope and wanted. - Running the service on a network you do not control, or exposing it to the internet. It is designed for two home networks and nothing else.
0.1.0, published on 2026-10-08, is the first release, 0.2.0,
published on 2026-10-11, the second, 0.3.0, published on 2026-10-15, the
third, 0.4.0, published on 2026-10-20, the fourth, 0.5.0, published on
2026-10-25, the fifth, and 0.6.0, published on 2026-10-30, the sixth.
Releases are published by the release workflow, from a tag, on this
repository's Releases page and nowhere else, signed under the name given
below.
| Version | What it is | Reports about it |
|---|---|---|
0.6.0 |
The sixth release, published on 2026-10-30. | Wanted, as described above, with no response-time commitment. A fix, if one is made, comes as a new release. |
0.5.0 |
The fifth release, published on 2026-10-25. It has four problems that 0.6.0 corrects, each found in the code and none seen to happen on a machine: a lookup takes a response sent by unicast, or from a port other than 5353, as its answer (REQ-RES-012), and believes an answer for as long as its TTL says, however long (REQ-RES-013); the conflict check on the client network reads such a response (REQ-ADV-030); and the probe keeps any DNS response that reaches it (REQ-SEC-018). README.md, "Known problems", says more. |
Wanted, as described above, with no response-time commitment. A fix, if one is made, comes as a new release. |
0.4.0 |
The fourth release, published on 2026-10-20. It has the four problems of 0.5.0 that 0.6.0 corrects. |
Wanted, as described above, with no response-time commitment. A fix, if one is made, comes as a new release. |
0.3.0 |
The third release, published on 2026-10-15. It has a fault that 0.4.0 corrects: a client connection accepted in the moment the service stops, or withdraws the printer, can be left open (REQ-LIF-002). It also has the four problems of 0.5.0 that 0.6.0 corrects. |
Wanted, as described above, with no response-time commitment. A fix, if one is made, comes as a new release. |
0.2.0 |
The second release, published on 2026-10-11. It has two faults that 0.3.0 corrects: it answers a legacy mDNS query by unicast to whatever address the query gives (REQ-SEC-017), and it does not stop cleanly when Windows shuts down (REQ-LIF-009); README.md, "Known problems", says more. It also has the fault of 0.3.0 that 0.4.0 corrects, and the four problems of 0.5.0 that 0.6.0 corrects. |
Wanted, as described above, with no response-time commitment. A fix, if one is made, comes as a new release. |
0.1.0 |
The first release, published on 2026-10-08. It has the same two faults as 0.2.0, which 0.3.0 corrects, the one 0.4.0 corrects, and the four problems 0.6.0 corrects. |
Wanted, as described above, with no response-time commitment. A fix, if one is made, comes as a new release. |
0.1.0-rc.1, 0.1.0-rc.2, 0.1.0-rc.3 |
Trials of the release workflow, and a pre-release for testing. Only 0.1.0-rc.3 was published. |
Not covered. 0.1.0 is compiled from the same source code as 0.1.0-rc.3; see whether the problem happens on 0.1.0. |
main |
The source as it stands. | No support promise of any kind. A gap between the specification and the code is in scope, as above. |
The .NET runtime each release carries is fixed when the release is built,
and a fix to it arrives only in a new release. 0.6.0 and 0.5.0 carry
10.0.14, from Microsoft's out-of-band security update of 2026-10-25, which was
still the latest .NET 10 release when 0.6.0 was built; 0.3.0 and 0.4.0
carry 10.0.13, and 0.1.0 and 0.2.0 carry 10.0.12. Microsoft's announcement
of that update lists four advisories, and says its update of 2026-10-13, which
brought 10.0.13, did not fully address two of them, one being CVE-2026-70197,
in System.Net.Sockets on Windows. Read against SecretPrinter on 2026-10-28,
by what Microsoft says each fault needs, SecretPrinter reaches none of the
four: it decompresses nothing (CVE-2026-111497, the Brotli decoder), never
hands one SocketAsyncEventArgs from an IPv4 socket to an IPv6 one
(CVE-2026-70197), is neither built nor installed from a Linux archive
(CVE-2026-114148), and uses no Blazor package (CVE-2026-102870). 0.1.0 to
0.4.0 still carry a runtime Microsoft lists as affected by the first two;
0.5.0 and 0.6.0 carry the corrected one, and moving to either settles both
without relying on that reading
(finding).
(Paragraph added 2026-10-25 by Claude, Anthropic model, Claude Opus 5.5, at the
direction of Edwin West, after release 0.5.0 was published. Changed 2026-10-28
by Claude, Claude Opus 5.5: until then it ended "Whether any of the four
concerns code SecretPrinter runs has not been worked out", with a link to the
finding of the fifth release. Changed 2026-10-30 by Claude, Claude Opus 5.5,
after release 0.6.0 was published: the second sentence began "0.5.0 carries
10.0.14, from Microsoft's out-of-band security update of 2026-10-25", and the
one on the corrected runtime said "0.5.0 carries the corrected one, and
moving to it settles both"; see
the finding of the sixth release.)
Consequently: a binary presented as SecretPrinter that does not come from this repository's Releases page, or is not signed under the name given below, does not come from this project, and you should not run it. The one exception is a build you made yourself from source.
(The table added, and the paragraph above it rewritten, 2026-10-08 by Claude,
Anthropic model, Claude Opus 5.5, at the direction of Edwin West, in the commit
that sets the version to 0.1.0. The paragraph had said that no release had been
published, that 0.1.0-rc.3 was a pre-release published on 2026-10-07 that
carried no support promise, and that neither did main; and the section ended
"This section will be replaced with a real version table at the first
release." Later the same day, once 0.1.0 had been published, "is the first
version meant as a release" became "published on 2026-10-08, is the first
release", in the first sentence and in the table.)
(The row for 0.2.0 added, and "and 0.2.0 is the second version meant as
one" added to the first sentence, 2026-10-11 by Claude, Anthropic model,
Claude Opus 5.5, at the direction of Edwin West, in the commit that sets the
version to 0.2.0. The row for 0.1.0 is unchanged. Later the same day, once
0.2.0 had been published, "is the second version meant as one" became
"published on 2026-10-11, the second", and "The second version meant as a
release." in the row became "The second release, published on 2026-10-11.")
(The row for 0.3.0 added, "and 0.2.0, published on 2026-10-11, the
second" changed to "0.2.0, published on 2026-10-11, the second, and 0.3.0
the third version meant as one" in the first sentence, and a sentence on the
two faults 0.3.0 corrects added to the rows for 0.2.0 and 0.1.0,
2026-10-15 by Claude, Anthropic model, Claude Opus 5.5, at the direction of
Edwin West, in the commit that sets the version to 0.3.0. Later the same day,
once 0.3.0 had been published, "the third version meant as one" became
"published on 2026-10-15, the third", and "The third version meant as a
release." in the row became "The third release, published on 2026-10-15.")
(The row for 0.4.0 added, "and 0.3.0, published on 2026-10-15, the third"
changed to "0.3.0, published on 2026-10-15, the third, and 0.4.0 the
fourth version meant as one" in the first sentence, and a sentence on the
fault 0.4.0 corrects added to the rows for 0.3.0, 0.2.0 and 0.1.0,
2026-10-20 by Claude, Anthropic model, Claude Opus 5.5, at the direction of
Edwin West, in the commit that sets the version to 0.4.0. Later the same day,
once 0.4.0 had been published, "the fourth version meant as one" became
"published on 2026-10-20, the fourth", and "The fourth version meant as a
release." in the row became "The fourth release, published on 2026-10-20.")
(The row for 0.5.0 added, and "and 0.4.0, published on 2026-10-20, the
fourth" changed to "0.4.0, published on 2026-10-20, the fourth, and 0.5.0
the fifth version meant as one" in the first sentence, 2026-10-25 by Claude,
Anthropic model, Claude Opus 5.5, at the direction of Edwin West, in the commit
that sets the version to 0.5.0. No sentence was added to the rows for earlier
versions. The one correction 0.5.0 carries, a lookup no longer taking a
goodbye record, one with a TTL of 0, as its answer, was found by reading the
code, and what it did in earlier versions has not been seen on a machine;
CHANGELOG.md describes it. Later the same day, once 0.5.0 had been
published, "the fifth version meant as one" became "published on 2026-10-25,
the fifth", and "The fifth version meant as a release." in the row became "The
fifth release, published on 2026-10-25.")
(The row for 0.6.0 added, "and 0.5.0, published on 2026-10-25, the fifth"
changed to "0.5.0, published on 2026-10-25, the fifth, and 0.6.0 the sixth
version meant as one" in the first sentence, and a sentence on the four
problems 0.6.0 corrects added to the rows for 0.5.0 to 0.1.0, 2026-10-30
by Claude, Anthropic model, Claude Opus 5.5, at the direction of Edwin West, in
the commit that sets the version to 0.6.0. Unlike the correction 0.5.0
carries, for which no sentence was added, each of the four has an entry in
"Known problems" in README.md naming the releases up to 0.5.0, so the rows
now say the same. Each was found in the code; none was seen to happen on a
machine. They are called problems, the word of "Known problems", and not
faults, because one of them, REQ-RES-013, departs from what RFC 6762
recommends, not from what it requires. Later the same day, once 0.6.0 had been
published, "the sixth version meant as one" became "published on 2026-10-30,
the sixth", and "The sixth version meant as a release." in the row became "The
sixth release, published on 2026-10-30.")
(The first two paragraphs replaced 2026-10-07 by Claude, Anthropic model, Claude Opus 5.5, at the direction of Edwin West, who decided that day to publish a pre-release. They had said that no release had been tagged, that the only way to run this was to build it from source, and that any binary presented as a SecretPrinter release did not come from this project because none existed. Later the same day, once the pre-release had been published, "If it has been published, it is on" became "It was published on 2026-10-07, on".)
Signed distribution on GitHub is the goal of this project, and it is written into the specification rather than left as an intention:
| Requirement | What it says |
|---|---|
REQ-DIST-004 |
Release binaries are signed with Azure Trusted Signing. |
REQ-DIST-005 |
The release documentation states how to verify a signature, and that unsigned binaries are not official releases. |
REQ-DIST-006 |
Every release is a tagged commit with a changelog entry. |
All three were unmet until 2026-10-04, and the specification checker
(tools/SecretPrinter.SpecCheck) failed the build because of it. That failure
was left visible on purpose: a project that asks to be trusted should not hide
the requirements it has not yet met.
On 2026-10-04 the release workflow signed a trial build, and the signatures
were read on a downloaded copy
(finding).
docs/verification.md records what stands behind each of the three
requirements, and what does not.
How to check a release is in
docs/operating.md:
compare the zip's SHA-256 with the published one, then read the signature on
every program file. SecretPrinter's own files are signed as:
CN=Edwin West, O=Edwin West, L=Huntington, S=ny, C=US
A SecretPrinter file that is unsigned, or signed under any other name, is not an official release.
The tags before v0.1.0 are not releases. v0.1.0-rc.1 and
v0.1.0-rc.2 were trials of the workflow, and nothing was published for
either. v0.1.0-rc.3 is the tag of a pre-release for testing. v0.1.0 is the
tag of the first version meant as a release, v0.2.0 of the second,
v0.3.0 of the third, v0.4.0 of the fourth, v0.5.0 of the fifth, and
v0.6.0 of the sixth; see "Supported versions" above.
(Sentence about v0.1.0-rc.3 added, and "The two tags that exist" reworded,
2026-10-07 by Claude, Claude Opus 5.5. Reworded again 2026-10-08, in the commit
that sets the version to 0.1.0: this paragraph began "There is still no
release.", and the sentence about v0.1.0 is new. "and v0.2.0 of the
second" added 2026-10-11 by Claude, Claude Opus 5.5, in the commit that sets
the version to 0.2.0. It became "v0.2.0 of the second, and v0.3.0 of the
third" on 2026-10-15, by Claude, Claude Opus 5.5, in the commit that sets the
version to 0.3.0. "and v0.4.0 of the fourth" added 2026-10-20 by Claude,
Claude Opus 5.5, in the commit that sets the version to 0.4.0. It became
"v0.4.0 of the fourth, and v0.5.0 of the fifth" on 2026-10-25, by Claude,
Claude Opus 5.5, in the commit that sets the version to 0.5.0, and
"v0.5.0 of the fifth, and v0.6.0 of the sixth" on 2026-10-30, by Claude,
Claude Opus 5.5, in the commit that sets the version to 0.6.0.)
(This section replaced 2026-10-04 by Claude, Anthropic model, Claude Opus 5.5, at the direction of Edwin West. From here down it had said that all three requirements were unmet, that verification instructions would be added when there was a signature to verify, and that every artifact should be treated as unsigned because every artifact was.)