Rendered at 02:40:19 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
crote 16 hours ago [-]
> With this approach, they have shown a reduction in CRL data from a list of all enrolled and unexpired certificate serial numbers from 6.7G to a filter of just 1.3 MB.
It actually gets even better: Mozilla's CRLite deltas with generation on a 6-hour interval are only about 60kB! Even with 10 billion internet devices fetching every 6 hours you're looking at a mere 222Gbps of bandwidth: a lot, but plenty of tech companies do more.
> At this point, why not just use DANE (RFC 6698), store the public keys in the DNS, rely on DNSSEC to provide the necessary authenticity, and use DNS TTL settings to control the cached lifetime of the public key? With a combination of DNSSEC Chain Extensions and DNSSEC stapling, it is possible to perform a security association handshake by having the server provide the client with (...)
Yeah, until an attacker gets access to DNS, stores a record with a TTL of 1 year, and staples that - of course after poisoning the caches of major DNS resolvers with the address of the attacker's server.
xorcist 15 hours ago [-]
> until an attacker gets access to DNS, stores a record with a TTL of 1 year,
DNSSEC may have problems, but that's not how the trust model works. Also signing is separate from authoritative DNS, so you'd need to compromise the signing itself, not just DNS. Should that happen, you are still limited by the upstream record siganture lifetime.
pjf 16 hours ago 14">[-]
On the other front (Chrome), their crlset-tools [1] just fetched me 64k (~1.1MiB) of revoked certs just fine, contrary to the article (quote: "After retrieving and running this tool, I was surprised to see a total of 1,081 revoked certificate serial numbers in this list. This seems oddly low.")
Interests of the existing PKI industry may be the source of some friction, but the bigger issue is that DANE depends on DNSSEC, which is not widely deployed, and sometimes actively avoided due to its complexity and ease of breaking you site.
Don't get me wrong, I'd love it if DANE, or something similar caught on, but I don't think it is practical until something changes to make DNSSEC (or equivalent) common.
jeroenhd 15 hours ago [-]
If DANE were to roll out to browsers, I think plenty of people would rather use it than centralizing on Let's Encrypt.
DNSSEC isn't easy, but either is certbot. DNSSEC also isn't that hard if you're not self-hosting your DNS servers (and even then it's easy if you pick a modern DNS server).
Most domains seem to use their registrars free DNS servers. For those domains, DNSSEC is often just a checkbox. I just activated DNSSEC on three domains by hitting that checkbox. A certbot-style tool can use the same API many existing certbot plugins already provide access to for setting up DANE.
However, until browsers actually implement DANE, it's pretty useless. I know some people use it for mail servers (for some reason, don't see why they can't use Let's Encrypt for that) but even there it's optional.
tptacek 3 hours ago [-]
The "just check a checkbox in your registrar" UX depends on the registrar having custody of your keys. That's not how certbot works.
thayne 9 hours ago [-]
> DNSSEC also isn't that hard if you're not self-hosting your DNS servers
It isn't that hard if:
- you use your domain registrar for serving your DNS as well. Even if you aren't "self-hosting", but use a different service for DNS hosting than you registered your domain with, then it can be complicated to coordinate between them. And
- your domain registrar makes it easy to set up. For some it is just a checkbox, for others, you have to contact customer support, and sometimes pay more.
Also, the server is just half of the problem. It also requires dns resolvers and clients to validate DNSSEC, which often isn't done today, and even when it is, often fails open, because so many domains don't use DNSSEC, and intermediate resolvers don't always support it. Validating DNSSEC can also hurt performance, in part because of the larger response sizes.
jeroenhd 7 hours ago [-]
> Even if you aren't "self-hosting", but use a different service for DNS hosting than you registered your domain with, then it can be complicated to coordinate between them
Indeed, you'd need to copy-paste four text fields
> for others, you have to contact customer support, and sometimes pay more.
That's ridiculous, I've never seen any registrar do that. Even if you do choose a terrible registrar, actual DANE rollout in browsers would put pressure on them to get their shit together.
Obviously, for DANE to work, verification must happen. DANE-enabled browsers will enforce validation, or fall back to regular TLS (with the scary warnings if someone stripped DNSSEC for a DANE server, as the certificate doesn't work any more). On operating systems that don't bother with DNSSEC validation, browsers can still query the necessary keys.
As for performance, DNSSEC does impose extra network traffic, but so does transmitting an intermediate certificate.
tptacek 3 hours ago [-]
Hard to square this with the operational history of DNSSEC at some of the best-resourced ops teams in the world.
PunchyHamster 18 hours ago [-]
> Interests of the existing PKI industry may be the source of some friction, but the bigger issue is that DANE depends on DNSSEC, which is not widely deployed, and sometimes actively avoided due to its complexity and ease of breaking you site.
I have a feeling it is "actively avoided" because vendors don't want to lose control of the cert ecosystem. Allowing user to just generate a domain for themselves means it will never get logged in central log and so can't be automatically found by crawlers by the big guys
Parodper 14 hours ago [-]
> Allowing user to just generate a domain for themselves
That's limited mostly by policy[1], the current PKI environment already allows delegating CA for a single domain.
Last time I checked support for that on client side was pretty spotty
tptacek 3 hours ago [-]
There is no support for DANE on the client side!
xorcist 15 hours ago [-]
This is public data so the big guys could absoltely crawl it. But we should not underestimate the size of the PKI industry, several large actors make good living from the existing web PKI and they will not change unless their very existence is threatened.
Parodper 14 hours ago [-]
It's funny to see that the issues with X.509 certificates, are being solved by what X.509 was intended to be used for: a directory system. It's DNS instead of X.500, but it's a start.
bblb 18 hours ago [-]
DNS and PKI. Two of the most centralized services in the Internet. Take over both of them, and you have the whole net under your command.
pjf 16 hours ago [-]
Good that at least BGP is secure.
nanis 16 hours ago [-]
Might want to add /sarc just in case someone believes it :-)
jumpconc 14 hours ago [-]
Just DNS. If you take over DNS, you can get Let's Encrypt to issue any certificate you want.
pjf 13 hours ago [-]
There are situations [1] where you could reliably BGP-hijack the IP prefix of the target domain authoritative nameserver, and obtain your own domain-validated cert for the target (by effectively controlling the zone file contents). And yeah, CAs do have their BGP protections, but still there's at least partial assumption BGP is secure enough to run DNS-based validation for new SSL certs, in our world where DNSSEC is still rare.
[1] https://www.ietf.org/proceedings/104/slides/slides-104-maprg-dns-observatory-monitoring-global-dns-for-performance-and-security-pawel-foremski-and-oliver-gasser-00.pdf (see slide 15; yeah, it's already a bit old, yet still the case from my practice)
PeterWhittaker 13 hours ago [-]
One quibble with the article: the notion that CRLs have to be large. When I was with Entrust our first releases targeted early Windows versions with limited memory, back when most Internet connections and even local networks were slow.
To ensure that RLs would always be manageable in size, we used distribution points (cRL and issuing) and decided at certificate issuance which RL would contain this certificate's serial number if ever it were revoked.
This approach scaled really well and kept RLs manageable.
There were applications that didn’t understand distribution points and needed the One RL to Revoke Them All, so we supported that as well (as an option, IIRC).
lmm 21 hours ago [-]
[flagged]
crote 16 hours ago [-]
> Revocation is an emergency measure, not a routine one. That's ok.
Rather the opposite: revocation is the one time it actually matters, so your infrastructure shouldn't come to a grinding halt when it happens!
Let's say a party like LetsEncrypt needs to do a mass revocation of all certificates. Unlikely, but it has happened before. This is going to instantly blow up the CRL from perhaps a few thousand to 700 million entries. Force every browser to download that regularly and you've essentially accidentally created a DDoS on LetsEncrypt's CRL service.
And how do you want the browser to respond? Fail-closed and you've just created a method to take 80% of the internet offline by DDoSing a single service, fail-open and you've just created a method for an attacker to bypass certificate revocation entirely.
With critical infrastructure like this you can't get away with only thinking about the happy path. It should always work - even in emergencies.
fc417fc802 14 hours ago [-]
> fail-open and you've just created a method for an attacker to bypass certificate revocation entirely
I generally agree about the rest but isn't what you suggest there sufficiently disproportionate to fall well outside the threat model? It buys only a limited window of opportunity in a very specific scenario while painting a neon target on your back. I feel like it's vaguely akin to worrying about a military checkpoint failing open when hit with a 30k lb bomb.
On revocation, check out https://bugzilla.mozilla.org/buglist.cgi?product=CA%20Progra...
I don't think any CA hasn't had an issue with revocation at some point (e.g. Let's Encrypt had a major one in 2021, and refused to revoke), which is why Let's Encrypt is moving to 7 day certs (so that revocation isn't required, basically https://www.imperialviolet.org/2011/03/18/revocation.html which is mentioned in the article). My impression is CRLs (and by implication current revocation methods) don't work, and browsers are effectively fudging around CAs with custom methods (e.g. allowing existing certs but no new certs from distrusted CAs).
I'm no security expert, but modern bind9 seems to just handle DNSSEC with no issues when I've used it, and given that the "WebPKI" seems is becoming more and more reliant on custom browser code, adopting DANE outside browsers might not be the worst idea.
lmm 17 hours ago [-]
> I don't think any CA hasn't had an issue with revocation at some point (e.g. Let's Encrypt had a major one in 2021, and refused to revoke)
Every software org has had issues with every piece of functionality, revocation isn't special.
> modern bind9 seems to just handle DNSSEC with no issues when I've used it
The happy path works. Everything is fine until it isn't. Very few people are confident enough to fully deploy it.
aragilar 16 hours ago [-]
According to https://stats.labs.apnic.net/dnssec DNSSEC is sitting about 1/3, so "very few" isn't accurate. I'm not suggesting browsers should change what they do, but if WebPKI can't be used, building a new CA ecosystem would seem to be to be at least as hard as getting DANE working.
jcgl 20 hours ago [-]
Agreeing with my sibling commenter, this writer is extremely experienced and has been writing this blog for many years. I’d be appalled if this were LLM-written, but thankfully it reads with the same style and tone that he’s always had.
> Revocation is an emergency measure, not a routine one. That's ok.
At the scale modern CAs operate, even emergency measures (i.e. measures that are an emergency for the party receiving the leaf cert) are also routine for the CA/the party granting the leaf cert.
PunchyHamster 18 hours ago [-]
> Because DNS' multilayered caching makes it notoriously impossible to operate safely or debug.
That is not a problem for certs, you are not changing it every second. And the "impossible to operate or debug" is just plain failse or incompetence
> Most large outages already originate in DNS issues; putting the crypto in that layer would redouble it.
That is also just not true. Also, outage of DNS coz someone fucked up configuration management somewhere is not caused by anything related to DNS, it just so happens DNS is essential so any problem is visible.
lmm 17 hours ago [-]
> That is not a problem for certs, you are not changing it every second.
The problem is when you screw it up and can't fix it for 24 hours or worse.
PunchyHamster 12 hours ago [-]
Well then don't set validity for 24 hours.
lmm 3 hours ago [-]
Most technologies work great when you simply don't make mistakes.
_bernd 20 hours ago [-]
The author has a long history of quality content and should be considered experienced in his field of doing.
miyuru 17 hours ago [-]
looks like so, the emojies are dead give away.
I am not against using LLMs, but the author should have validated the contents before posting.
It actually gets even better: Mozilla's CRLite deltas with generation on a 6-hour interval are only about 60kB! Even with 10 billion internet devices fetching every 6 hours you're looking at a mere 222Gbps of bandwidth: a lot, but plenty of tech companies do more.
> At this point, why not just use DANE (RFC 6698), store the public keys in the DNS, rely on DNSSEC to provide the necessary authenticity, and use DNS TTL settings to control the cached lifetime of the public key? With a combination of DNSSEC Chain Extensions and DNSSEC stapling, it is possible to perform a security association handshake by having the server provide the client with (...)
Yeah, until an attacker gets access to DNS, stores a record with a TTL of 1 year, and staples that - of course after poisoning the caches of major DNS resolvers with the address of the attacker's server.
DNSSEC may have problems, but that's not how the trust model works. Also signing is separate from authoritative DNS, so you'd need to compromise the signing itself, not just DNS. Should that happen, you are still limited by the upstream record siganture lifetime.
[1] https://github.com/agl/crlset-tools
Interests of the existing PKI industry may be the source of some friction, but the bigger issue is that DANE depends on DNSSEC, which is not widely deployed, and sometimes actively avoided due to its complexity and ease of breaking you site.
Don't get me wrong, I'd love it if DANE, or something similar caught on, but I don't think it is practical until something changes to make DNSSEC (or equivalent) common.
DNSSEC isn't easy, but either is certbot. DNSSEC also isn't that hard if you're not self-hosting your DNS servers (and even then it's easy if you pick a modern DNS server).
Most domains seem to use their registrars free DNS servers. For those domains, DNSSEC is often just a checkbox. I just activated DNSSEC on three domains by hitting that checkbox. A certbot-style tool can use the same API many existing certbot plugins already provide access to for setting up DANE.
However, until browsers actually implement DANE, it's pretty useless. I know some people use it for mail servers (for some reason, don't see why they can't use Let's Encrypt for that) but even there it's optional.
It isn't that hard if:
- you use your domain registrar for serving your DNS as well. Even if you aren't "self-hosting", but use a different service for DNS hosting than you registered your domain with, then it can be complicated to coordinate between them. And
- your domain registrar makes it easy to set up. For some it is just a checkbox, for others, you have to contact customer support, and sometimes pay more.
Also, the server is just half of the problem. It also requires dns resolvers and clients to validate DNSSEC, which often isn't done today, and even when it is, often fails open, because so many domains don't use DNSSEC, and intermediate resolvers don't always support it. Validating DNSSEC can also hurt performance, in part because of the larger response sizes.
Indeed, you'd need to copy-paste four text fields
> for others, you have to contact customer support, and sometimes pay more.
That's ridiculous, I've never seen any registrar do that. Even if you do choose a terrible registrar, actual DANE rollout in browsers would put pressure on them to get their shit together.
As for DNSSEC validation: validation currently seems to happen between 0 to 95% according to https://stats.labs.apnic.net/dnssec
Obviously, for DANE to work, verification must happen. DANE-enabled browsers will enforce validation, or fall back to regular TLS (with the scary warnings if someone stripped DNSSEC for a DANE server, as the certificate doesn't work any more). On operating systems that don't bother with DNSSEC validation, browsers can still query the necessary keys.
As for performance, DNSSEC does impose extra network traffic, but so does transmitting an intermediate certificate.
I have a feeling it is "actively avoided" because vendors don't want to lose control of the cert ecosystem. Allowing user to just generate a domain for themselves means it will never get logged in central log and so can't be automatically found by crawlers by the big guys
That's limited mostly by policy[1], the current PKI environment already allows delegating CA for a single domain.
[1] https://community.letsencrypt.org/t/sub-ca-with-wildcard-cer...
To ensure that RLs would always be manageable in size, we used distribution points (cRL and issuing) and decided at certificate issuance which RL would contain this certificate's serial number if ever it were revoked.
This approach scaled really well and kept RLs manageable.
There were applications that didn’t understand distribution points and needed the One RL to Revoke Them All, so we supported that as well (as an option, IIRC).
Rather the opposite: revocation is the one time it actually matters, so your infrastructure shouldn't come to a grinding halt when it happens!
Let's say a party like LetsEncrypt needs to do a mass revocation of all certificates. Unlikely, but it has happened before. This is going to instantly blow up the CRL from perhaps a few thousand to 700 million entries. Force every browser to download that regularly and you've essentially accidentally created a DDoS on LetsEncrypt's CRL service.
And how do you want the browser to respond? Fail-closed and you've just created a method to take 80% of the internet offline by DDoSing a single service, fail-open and you've just created a method for an attacker to bypass certificate revocation entirely.
With critical infrastructure like this you can't get away with only thinking about the happy path. It should always work - even in emergencies.
I generally agree about the rest but isn't what you suggest there sufficiently disproportionate to fall well outside the threat model? It buys only a limited window of opportunity in a very specific scenario while painting a neon target on your back. I feel like it's vaguely akin to worrying about a military checkpoint failing open when hit with a 30k lb bomb.
On revocation, check out https://bugzilla.mozilla.org/buglist.cgi?product=CA%20Progra... I don't think any CA hasn't had an issue with revocation at some point (e.g. Let's Encrypt had a major one in 2021, and refused to revoke), which is why Let's Encrypt is moving to 7 day certs (so that revocation isn't required, basically https://www.imperialviolet.org/2011/03/18/revocation.html which is mentioned in the article). My impression is CRLs (and by implication current revocation methods) don't work, and browsers are effectively fudging around CAs with custom methods (e.g. allowing existing certs but no new certs from distrusted CAs).
I'm no security expert, but modern bind9 seems to just handle DNSSEC with no issues when I've used it, and given that the "WebPKI" seems is becoming more and more reliant on custom browser code, adopting DANE outside browsers might not be the worst idea.
Every software org has had issues with every piece of functionality, revocation isn't special.
> modern bind9 seems to just handle DNSSEC with no issues when I've used it
The happy path works. Everything is fine until it isn't. Very few people are confident enough to fully deploy it.
> Revocation is an emergency measure, not a routine one. That's ok.
At the scale modern CAs operate, even emergency measures (i.e. measures that are an emergency for the party receiving the leaf cert) are also routine for the CA/the party granting the leaf cert.
That is not a problem for certs, you are not changing it every second. And the "impossible to operate or debug" is just plain failse or incompetence
> Most large outages already originate in DNS issues; putting the crypto in that layer would redouble it.
That is also just not true. Also, outage of DNS coz someone fucked up configuration management somewhere is not caused by anything related to DNS, it just so happens DNS is essential so any problem is visible.
The problem is when you screw it up and can't fix it for 24 hours or worse.
I am not against using LLMs, but the author should have validated the contents before posting.