If you don't have an expiry then you'd need to keep certificate revocations around forever, so your CRL grows without bound. Currently the certificate revocation list can remove expired certificates as clients won't use them. There's also lots of clients that don't check CRLs. Those clients would be forever vulnerable to a stolen certificate if we removed expiry.
Imagine if you inherited the infrastructure for a company where the previous admin created certs, lost track of them, and used questionable methods for securing them. I believe you'd struggle to find and revoke every previously issued certificate. The expiry ensures that gets cleaned up, eventually.
I don't have a good solution to the Nokia Symbian problem. I suspect if the device only supports SHA-1 then there are other unpatched security issues. Does anyone know if that's the case? If so, this seems more like a support and patching issue.
I told that mostly as an anecdote. I don't think a solution is really expected. The devices I referred to (made 2005 - 2007) worked rather well for about 10 years. I guess that's a satisfactory achievement.
This is in part due to the difficulty of doing so. On linux at least there is generally not a central location for CRLs, or anything to keep them up to date. Each application is responsible for maintinging CRLs it cares about itself. And then there is the fact that a full set of CRLs is pretty big. OCSP fixes some of that, at the cost of more latency during the TLS handshake. OCSP stapling is probably the best existing solution, but I don't think it is widespread enough that you can rely on it yet.
(Disclaimer: I work for Mozilla and have helped with the rollout of this)
There are multiple reasons, but you seem focused on the certificate management side. One advantage of doing things more frequently is they are forced to become more routine or automated. So shorter expirations should actually make expiration failures less likely over time. Of course the transition sucks.
That's a meaningful distinction. Certificates which lack SCTs (the proof that they were shown to Certificate Transparency logs before you) don't work in some popular browsers, but those certificates are not policy violations, they just don't work in browsers so you probably should not use them on a web site. In a handful of cases such certificates exist for legacy reasons (e.g. industrial environment that doesn't know anything about the "Web"), in other cases they're minted but not intended to be seen yet, for example Google's front end facing systems can do this.
When Google accidentally made some of those certificates live with insufficient SCTs they just did not work in Chrome - which is embarrassing but it was not a policy violation, the subsequent root cause analysis was Google's choice not mandated by the other root trust stores and there was no threat that anybody's trust would get revoked.
* CRL is a list of revoked certificates, it must be downloaded and manually applied to all clients, which only occurs once in a while as a new update, and many hate system updates. The huge lists of revoked certificate have size, performance and scalability problems, currently they are considered obsolete and mainly replaced by OCSP.
* In OCSP, clients ask a CA's OCSP server to determine the validity of a certificate. It operates out-of-band from the main TLS connection, and networks and CA servers are not reliable, especially in the early days. Timed out OCSP requests were common. If you are a web browser, you don't want a failed OCSP request to block/DoS HTTPS, thus OCSP becomes essentially "optional". Browsers reject bad certificates when the connection goes through, but if it times out, nothing happens. An attacker can simply block the OCSP server to bypass the revocation check. There's also the problem of privacy - you have to send the name of every website you visit to third party servers.
* In OCSP Stapling, the webserver acts like a proxy - it hosts a cached copy of the OCSP response for itself (it cannot be forged because it's signed by the CA) and sends it to clients in-band, via normal TLS channel. Because now the client no longer asks CA's OCSP servers for a response, but instead simply asks the webserver itself, it eliminates the problem of third-party OCSP servers, solves the reliability and privacy issues. However, there's nothing to prevent an attacker to run a server with a compromised certificate with OCSP Stapling turned off.
* In OCSP Must-Staple, you can get a certificate from a CA that says "you must use OCSP Stapling, otherwise this certificate is null and void." Hence, if a browser sees a certificate with "OCSP Must-Staple" enabled, it must see whether the webserver supports OCSP Stapling, and via OCSP Stapling, the certificate's validity will be determined. If OCSP Stapling is not supported by the webserver, the connection is rejected. Thus, an attacker who uses a compromised certificate has to either disable OCSP Stapling and be rejected, or to enable OCSP Stapling and gets caught. Finally the problem of reliable revocation is solved. But this is an optional feature and is only used by a tiny percentage of sysadmins. Also, it's only supported by webservers and browsers, other TLS-based applications like VPN, FTPS, SMTP, IMAP, XMPP, IRC servers/clients, etc, usually don't support OCSP Stapling at all.
Thus, it's often practically impossible to reliably revoke a certificate in SSL/TLS after someone has compromised your private key. The only feasible workaround is early expiration to reduce the window of vulnerability, an unrevokable certificate that expires within 90 days is less dangerous than a zombie certificate that lasts two years.
> In order to indicate their desire to receive certificate status information, clients MAY include an extension of type "status_request" in the (extended) client hello. Servers that receive a client hello containing the "status_request" extension MAY return a suitable certificate status response to the client along with their certificate.
And in a TLS handshake, the procedure is ClientHello, ServerHello, Server Certificate, CertificateRequest, Client Certificate. But "Certificate Status Request" is only specified in ClientHello, not CertificateRequest, so it's not possible for a server to request OCSP status from the client.
Hence, OCSP Stapling strictly a service provided to the clients, not vice versa - unfortunately, reliable revocation of client certificate is a still a problem. But I guess it's much easier to blacklist the offended client certificates on your own server, perhaps still problematic in a big organization, but certainly easier than blacklisting every leaked certificate on the Internet.
Disclamer: I don't have any experience on managing systems that use client verification. I did my best to RTMF and fact-check myself, but corrections are welcomed.
For example take a 90 day Let's Encrypt certificate. When it was issued ISRG had recently (likely seconds earlier) verified that the subscriber apparently had control over all the names listed.
Suppose a month later one of the names has been sold to you. You're entitled to have that certificate revoked. Since you control the name, you can just do the Let's Encrypt proof-of-control dance and get it revoked, issue a new one instead, whatever you want.
They issue millions of certificates every day, so they don't want to remain indefinitely responsible for the ones from last July, from 2018, and so on. Ninety days later, having expired, ISRG has no opinion about whether the holder of the certificate still controls those names, and if you call that revocation API it will refuse, there isn't anything to revoke because it has expired.
Now, if the certificate you're shown is 15 seconds expired, is it so much more likely that it's really because they don't control this name any more than it was a minute earlier? Not really, but we must draw a line in the sand somewhere, or else a 10 year old certificate for "www.microsoft.com" is still valid right?
For the root the argument is a bit trickier, after all they could just issue themselves a new certificate with the same keys but a different expiry date - and some root CAs have done this in the past.
In this case (for roots) as we see with Android not every system actually enforces expiry. Out of the box OpenSSL doesn't for example I think. But the argument as to why expiry matters for roots goes like this: If we are actively managing the trust store, clearly a trusted CA which wants us to continue trusting them should send us a new or updated root in plenty of time. If they forgot we can hardly have confidence in their other processes. If they decided not to we certainly should cease to trust their root when it expires as this signifies they presumably do not intend to obey policy any more, for this root at least.
For example, if you have a Firefox that you just decide not to ever update, I believe it won't enforce expiry of roots. Newer Firefox builds will, like clockwork, remove expired roots entirely. But the old build will continue to trust roots that have notionally expired if you are still running it.
This is the key issue. For a long time there was no working revocation mechanism. With a combination of OCSP, OCSP stapling and the muststaple extension you can get there, but this isn't in widespread use yet. Revocation is a difficult problem.
Having such a tempting footgun is a real problem.
I don't see a reason why I would ever want to use auth=none for anything as at that point JWT is just base64 encoded json but I may be too deep in the identity world by now.
Also, unrelated to that argument but I think the most expensive part in a CA is not issuance but rather OCSP since an OCSP query could happen with every TLS connection (although stapling exists to solve that). OCSP would not be affected by a shortening of validity periods.
It’s well-documented NSA saboteurs infiltrated the standards board. They forced through a bunch of bad proposals with the intention of making it overly complicated. The idea was to encourage misconfiguration and implementation bugs.
If anything, now that everything is connected to the internet you want shorter revocations (like days, weeks or months). That way the potential for abuse is shorter and the path for renewal is better trodden by organizations (ie less likely to forget about an expiring cert).
[1] https://www.theverge.com/2013/12/20/5231006/nsa-paid-10-mill...
[2] https://en.m.wikipedia.org/wiki/RSA_BSAFE
[3] https://en.m.wikipedia.org/wiki/Bullrun_(decryption_program)
https://www.mail-archive.com/[email protected]/msg12...
It’s describing the same tactics, but a different protocol. Honestly, just crack open the SSL spec. In hindsight, it’s pretty obvious it was intentionally over-complicated.
The Wireguard protocol attempts to fix these issues by hardcoding everything behind a protocol version number. It’s vastly easier to implement and configure properly.
If you have a security incident you can't wait for the expiration date anyway, you need to revoke. And hopefully users have a mechanism to note the revocation in time.
So expiration dates help to avoid using weak algorithms forever. But does that need to be done on a fixed date per certificate? Wouldn't that be a more gradual approach in the libraries? And of course they are not updated they get less secure as time goes by. And sooner or later compatibility issues will do the job. E. g. most Nokia Symbian phones don't do https any more, because all they can handle is SHA-1 certificate signatures. That was a gradual decline in functionality.