[jmileham]

I don't think he'd refute that he'd have more to worry about had he not been lucky enough to be on an unaffected version of OpenSSL. I believe his point was that the use of stunnel to terminate SSL connections mitigates some of the attack vectors that could've been used to recover customer information in the event of having been compromised at the OpenSSL layer, and that the architecture of Tarsnap itself absolutely precludes recovery of customer backups in any event. And that these facts aren't an accident.

The important takeaway from this post is that it pays to employ layers of security when building software systems.

[rst]

I was reacting mostly to the apparent exclusion of usernames, passwords, and session cookies (all exposed in net traffic) from the category of "anything sensitive".

[cperciva]

Fair point. Perhaps I should have said that the stunnel/jail setup keeps OpenSSL bugs away from the more sensitive things.

[ams6110]

If I understand correctly, he's decoupled the SSL connection handling from the http server. That seems to have all kinds of advantages to me, for example if a vulnerability in your SSL library is found you could quickly swap in SSL termination based on another library (e.g. gnutls or nss) or if there was a vulerablity in stunnel you could change it out for stud or even apache or nginx in a jail (assuming you had any or all of these things ready to go). Should also make for more flexibility with load balancing. Brilliant engineering.

[weavie]

Yes. And assume there will be bugs.