On 29.03.2022 11:44, rsbecker@xxxxxxxxxxxxx wrote:
On March 16, 2022 12:45 PM, I wrote:
On March 16, 2022 12:27 PM, Fabian Stelzer wrote:
On 16.03.2022 10:34, rsbecker@xxxxxxxxxxxxx wrote:
Following up on our IRC discussion on Monday, I have had a request to
support signing git commits and tags with SSL certificates instead of
SSH/GPG. The organization is heavily invested in SSL infrastructure,
so they want to go down that path.
The basic technique for doing this is, for example:
openssl dgst -sha256 -sign key -out content.sha256 signature.txt
-passin passphrase
There is a pre-step to compute the sha256, in this example, into a
file provided to openssl. We could use openssl to compute the hash also.
Verification is a bit different than what SSH or GPG does:
openssl dgst -sha256 -verify <(openssl x509 -in certificate -pubkey
-noout) -signature sign.txt.sha256 signature.txt
and reports either
Verified OK
Or
Verification Failure
It does not look like completion codes are consistently involved.
This also does look structurally different than both GPG and SSH and
more work to set up. It may be possible to provide wrappers and
pretend we are in SSH, but I'm not sure that is the right path to take.
Any pointers on how this might be done in existing git infrastructure,
or should I look into making this work in code? Sorry to say that the
documentation is not that clear on this.
Why not gpgsm? It can deal with x509 certs and is already supported. I
am using this to do s/mime signing/encryption with an yubikey hardware
token but static certs/keys should be even simpler. However I'm not
sure how good this works on other platforms.
Take a look into the GPGSM prereq in t/lib-gpg.sh for a few hints on
how to set this up.
Good idea but this is a non-starter. I have a limit of GPG 1.4, which only has the
single legacy object. GPG added a dependency to mmap, which is not available on
any of my platforms. That was one reason we were so happy to have SSH support.
I have been investigating this capability in more depth. After discussing with OpenSSL, explicitly adding SSL signing to git would introduce CVE-2022-0778 into git and allow a hostile upstream repo to introduce a deliberately defective key that could trigger this CVE unless customers have patched OpenSSL. Given the lack of broad-based adoption of the fixes to this point, I am reluctant to pursue this capability at this time. (Actually referencing my own advice in Git Rev News 82). The impact on git would be looping processes when signatures are evaluated. This would break workflows that depend on signed content and have downloaded keys with the CVE attributes.
Does anyone agree/disagree with me on delaying this?
--Randall
Do you actually need SSL Signing so you can verify commits with a single CA
key? Or do you have all the certs public keys anyway?
I know quite a few setups where every employee is issued an x509 cert (often
PIV Certs, preferably on a smartcard/token) and a central ldap is available
with all issued certs. This is usually used for authentication and s/mime.
However this can easily be used with ssh signing as well. I do so myself. I
use my own s/mime cert loaded into an ssh-agent (pkcs11 smartcard) to sign
commits and generate an allowed signers file with all the pubkeys extracted
from the certs i get from the PKIs ldap server.
