Am 14.11.2026 um 07:36 schrieb [email protected]:
> Hello misc@,
> 
> I'd like some critical eyes on aimap, an IMAP4rev2 (RFC 9051) server
> for OpenBSD. To say it first: all of its code was generated by
> Anthropic's Claude, under my direction. I designed it with the model,
> reviewed it and tested it, but I did not write the C.

It took me 10 seconds to copy and paste your mail unchanged into an AI
prompt which responded in 29 seconds with the following which in turn
took me longer to read, than I wanted to:

"Thanks for the transparency. The process model you describe is
recognizably in the style of smtpd(8)/relayd(8): root parent,
per-connection listener (chroot /var/empty, tight pledge), per-account
store (chroot to the maildir, drop to the account uid, unveil), a
separate parser process that never sees the spool directly, and a keymgr
holding the TLS private key. Using the existing
tls_config_use_fake_private_key() + ex_data pattern is consistent with
how smtpd and relayd already keep the private key out of the process
that talks to the network. On paper those choices are the right ones for
an IMAP daemon on OpenBSD.

That is also where the useful discussion ends for this project.

The entire implementation was produced by Claude. You designed the
architecture, directed the model, reviewed the output, and tested it,
but you did not write the C. Under the copyright regime that actually
exists today, that means:

You cannot grant OpenBSD (or anyone else) the rights it needs to
redistribute the code, because pure LLM output is not recognized as
copyrightable by the person who merely prompted it.
The model itself (and Anthropic) cannot grant those rights either.
There is no clean way to know what the training corpus contributed. For
a protocol as complex as IMAP4rev2, with MIME parsing, SEARCH,
CONDSTORE/QRESYNC, etc., that is not a theoretical concern.
This is the same problem that killed the earlier LLM-generated ext4
attempt. Theo’s statement then still applies: the chances of accepting
code with that copyright situation are zero. The fact that you are not
proposing this for base or for the ports tree does not remove the
problem for anyone who might later want to use, package, or audit it.
“Claims no copyright of its own” is not a license; it is an admission
that no usable license can be offered.

Beyond the legal issue there is a practical one. An IMAP server’s attack
surface is large (untrusted network input, MIME structure, search
queries, concurrent access to maildirs, UIDVALIDITY/modseq handling,
etc.). Privilege separation helps, but the code that implements the
protocol still has to be correct. When the only “author” is a model that
was trained on a mixture of existing IMAP implementations, the usual
ways we gain confidence—reading the code with the person who wrote it,
asking why a particular edge case was handled that way, knowing which
invariants the author was deliberately preserving—are weaker. Reviewing
the result is closer to reverse-engineering than to ordinary code
review. You have already noted the absence of independent human security
review; that is not a minor caveat for this class of program.

The keymgr approach is fine as a local design choice. There is not
currently a cleaner libtls API for “this process will never see the
private key,” which is why smtpd and relayd use the fake-key + ex_data
method. That does not change the larger problem.

If the goal is simply an experiment, that is your choice. Do not put it
on any machine that holds real mail or faces the network until it has
been rewritten by a human who understands every line and can stand
behind the copyright and the security properties. For the project’s
purposes, pure LLM-generated code is not something we will review for
inclusion, package, or recommend.

The design sketch is interesting. The provenance is not."



Reply via email to