TOFU recommendations for Gemini September 21, 2020 on Drew DeVault's blog

I will have more to say about Gemini in the future, but for now, I wanted to write up some details about one thing in particular: the trust-on-first-use algorithm I implemented for my client, gmni. I think you should implement this algorithm, too!

First of all, it’s important to note that the Gemini specification explicitly mentions TOFU and the role of self-signed certificates: they are the norm in Geminiland, and if your client does not support them then you’re going to be unable to browse many sites. However, the exact details are left up to the implementation. Here’s what mine does:

First, on startup, it finds the known_hosts file. For my client, this is ~/.local/share/gmni/known_hosts (the exact path is adjusted as necessary per the XDG basedirs specification). Each line of this file represents a known host, and each host has four fields separated by spaces, in this order:

• Hostname (e.g. gemini.circumlunar.space)
• Fingerprint algorithm (e.g. SHA-512)
• Fingerprint, in hexadecimal, with ‘:’ between each octet (e.g. 55:01:D8…)
• Unix timestamp of the certificate’s notAfter date

If a known_hosts entry is encountered with a hashing algorithm you don’t understand, it is disregarded.

Then, when processing a request and deciding whether or not to trust its certificate, take the following steps:

1. Verify that the certificate makes sense. Check the notBefore and notAfter dates against the current time, and check that the hostname is correct (including wildcards). Apply any other scrutiny you want, like enforcing a good hash algorithm or an upper limit on the expiration date. If these checks do not pass, the trust state is INVALID, GOTO 5.
2. Compute the certificate’s fingerprint. Use the entire certificate (in OpenSSL terms, X509_digest will do this), not just the public key.¹
3. Look up the known_hosts record for this hostname. If one is found, but the record is expired, disregard it. If one is found, and the fingerprint does not match, the trust state is UNTRUSTED, GOTO 5. Otherwise, the trust state is TRUSTED. GOTO 7.
4. The trust state is UNKNOWN. GOTO 5.
5. Display information about the certficate and its trust state to the user, and prompt them to choose an action, from the following options:
   • If INVALID, the user’s choices are ABORT or TRUST_TEMPORARY.
   • If UNKNOWN, the user’s choices are ABORT, TRUST_TEMPORARY, or TRUST_ALWAYS.
   • If UNTRUSTED, abort the request and display a diagnostic message. The user must manually edit the known_hosts file to correct the issue.
6. Complete the requested action:
   • If ABORT, terminate the request.
   • If TRUST_TEMPORARY, update the session’s list of known hosts.
   • If TRUST_ALWAYS, append a record to the known_hosts file and update the session’s list of known hosts.
7. Allow the request to proceed.

If the trust state is UNKNOWN, instead of requring user input to proceed, the implementation MAY proceed with the request IF the UI displays that a new certificate was trusted and provides a means to review the certificate and revoke that trust.

Note that being signed by a certificate authority in the system trust store is not considered meaningful to this algorithm. Such a cert is TOFU’d all the same.

That’s it! If you have feedback on this approach, please send me an email.

My implementation doesn’t entirely match this behavior, but it’s close and I’ll finish it up before 1.0. If you want to read the code, here it is.

Bonus recommendation for servers: you should use a self-signed certificate, and you should not use a certificate signed by one of the mainstream certificate authorities. We don’t need to carry along the legacy CA cabal into our brave new Gemini future.