As already noted on this thread, you can't use certbot today to get an IP address certificate. You can use lego [1], but figuring out the exact command line took me some effort yesterday. Here's what worked for me:
lego --domains 206.189.27.68 --accept-tos --http --disable-cn run --profile shortlived
It's a huge ask, but i'm hoping they'll implement code-signing certs some day, even if they charge for it. It would be nice if appstores then accepted those certs instead of directly requiring developer verification.
I wonder if transport mode IPsec can be relevant again if we're going to have IP address certificates. Ditto RFC 5660 (which -full disclosure- I authored).
Next, I hope they focus on issuing certificates for .onion addresses. On the modern web many features and protocols are locked behind HTTPS. The owner of a .onion has a key pair for it, so proving ownership is more trustworthy than even DNS.
I have now implemented a 2 week renewal interval to test the change to the 45 days, and now they come with a 6-day certificate?
This is no criticism, I like what they do, but how am I supposed to do renewals? If something goes wrong, like the pipeline triggering certbot goes wrong, I won't have time to fix this. So I'd be at a two day renewal with a 4 day "debugging" window.
I'm certain there are some who need this, but it's not me. Also the rationale is a bit odd:
> IP address certificates must be short-lived certificates, a decision we made because IP addresses are more transient than domain names, so validating more frequently is important.
Are IP addresses more transient than a domain within a 45 day window? The static IPs you get when you rent a vps, they're not transient.
The push for shorter and shorter cert lifetimes is a really poor idea, and indicates that the people working on these initiatives have no idea how things are done in the wider world.
For better or worse the push down to 47-day certificates is an industry-wide thing, in a few years nobody will be issuing certificates for longer than that.
Nobody is being forced to use 6-day certs though, when the time comes LE will default to 47 days just like everyone else.
Though if I may put on my tinfoil hat for a moment, I wonder if current algorithms for certificate signing have been broken by some government agency or hacker group and now they're able to generate valid certificates.
But I guess if that were true, then shorter cert lives wouldn't save you.
> broken by some government agency or hacker group
Probably not. For browsers to accept this certificate it has to be logged in a certificate transparency log for anyone to see, and no such certificates have been seen to be logged.
My browser on my work laptop has 219 root certificates trusted. Some of those may be installed from my employer, but I suspect most of them come from MS as it's Edge on Windows 11. I see in that list things like "Swedish Government Root Authority" "Thailand National Root Certification Authority" "Staat der Nederlanden Root CA" and things like "MULTICERT Root Certification Authority" "ACCVRAUZ1". I don't think there is any reason to believe any certificate. If a government wants a cert for a given DNS they will get it, either because they directly control a trusted root CA, or because they will present a warrant to a company that wants to do business in their jurisdiction and said company will issue the cert.
TLS certs should be treated much more akin to SSH host keys in the known hosts file. Browsers should record the cert the first time they see it and then warn me if it changes before it's expiration date, or some time near the expiration date.
Certificate transparency effectively means that any government actually uses a false certificate on the wider web and their root cert will get revoked.
Obviously you might still be victim #1 of such a scheme... But in general the CA's now aren't really trusted anymore - the real root of trust is the CT logs.
It's less about IP address transience, and more about IP address control. Rarely does the operator of a website or service control the IP address. It's to limit the CA's risk.
The short-lived requirement seems pretty reasonable for IP certs as IP addresses are often rented and may bounce between users quickly. For example if you buy a VM on a cloud provider, as soon as you release that VM or IP it may be given to another customer. Now you have a valid certificate for that IP.
6 days actually seems like a long time for this situation!
> Are IP addresses more transient than a domain within a 45 day window?
If I don't assign an EIP to my EC2 instance and shut it down, I'm nearly guaranteed to get a different IP when I start it again, even if I start it within seconds of shutdown completing.
It'd be quite a challenge to use this behavior maliciously, though. You'd have to get assigned an IP that someone else was using recently, and the person using that IP would need to have also been using TLS with either an IP address certificate or with certificate verification disabled.
If you are doing this in a commercial context and the 4 day debugging window, or any downtime, would cause you more costs than say, buying a 1 year certificate from a commercial supplier, then that might be your answer there...
Some ACME clients that I think currently support IP addresses are acme.sh, lego, traefik, acmez, caddy, and cert-manager. Certbot support should hopefully land pretty soon.
cert-manager maintainter chiming in to say that yes, cert-manager should support IP address certs - if anyone finds any bugs, we'd love to hear from you!
We also support ACME profiles (required for short lived certs) as of v1.18 which is our oldest currently supported[1] version.
We've got some basic docs[2] available. Profiles are set on a per-issuer basis, so it's easy to have two separate ACME issuers, one issuing longer lived certs and one issuing shorter, allowing for a gradual migration to shorter certs.
IP addresses must be accessible from the internet, so still no way to support TLS for LAN devices without manual setup or angering security researchers.
Also I don't see the point of what TLS is supposed to solve here? If you and I (and everyone else) can legitimately get a certificate for 10.0.0.1, then what are you proving exactly over using a self-signed cert?
There would be no way of determining that I can connecting to my-organisation's 10.0.0.1 and not bad-org's 10.0.0.1.
With a 6 day lifetime you'd typically renew after 3 days. If Lets Encrypt is down or refuses to issue then you'd have to choose a different provider. Your browser trusts many different "top of the chain" providers.
With a 30 day cert with renewal 10-15 days in advance that gives you breathing room
Personally I think 3 days is far too short unless you have your automation pulling from two different suppliers.
Thank you, I missed the part with several "top of the chain" providers. So all of them would need to go down at the same time for things to really stop working.
How many "top of chain" providers is letsencrypt using? Are they a single point of failure in that regard?
I'd imagine that other "top of chain" providers want money for their certificates and that they might have a manual process which is slower than letsencrypt?
But in general, one of the points of ACME is to eliminate dependence on a single provider, and prevent vendor lock-in. ACME clients should ideally support multiple ACME CAs.
For example, Caddy defaults to both LE and ZeroSSL. Users can additionally configure other CAs like Google Trust Services.
“Are they a single point of failure in that regard?”
It depends. If the ACME client is configured to only use Let’s Encrypt, then the answer is yes. But the client could fall-back to Google’s CA, ZeroSSL, etc. And then there is no single point of failure.
This is interesting, I am guessing the use case for ip address certs is so your ephemeral services can do TLS communication, but now you don't need to depend on provisioning a record on the name server as well for something that you might be start hundreds or thousands of, that will only last for like an hour or day.
One thing this can be useful for is encrypted client hello (ECH), the way TLS/HTTPS can be used without disclosing the server name to any listening devices (standard SNI names are transmitted in plaintext).
To use it, you need a valid certificate for the connection to the server which has a hostname that does get broadcast in readable form. For companies like Cloudflare, Azure, and Google, this isn't really an issue, because they can just use the name of their proxies.
For smaller sites, often not hosting more than one or two domains, there is hardly a non-distinct hostname available.
With IP certificates, the outer TLS connection can just use the IP address in its readable SNI field and encrypt the actual hostname for the real connection. You no longer need to be a third party proxying other people's content for ECH to have a useful effect.
Even if it did work, the privacy value of hiding the SNI is pretty minimal for an IP address that hosts only a couple domains, as there are plenty of databases that let you look up an IP address to determine what domain names point there - e.g. https://bgp.tools/prefix/18.220.0.0/14#dns
> In verifying the client-facing server certificate, the client MUST interpret the public name as a DNS-based reference identity [RFC6125]. Clients that incorporate DNS names and IP addresses into the same syntax (e.g. Section 7.4 of [RFC3986] and [WHATWG-IPV4]) MUST reject names that would be interpreted as IPv4 addresses.
I don't really see the value in ECH for self-hosted sites regardless. It works for Cloudflare and similar because they have millions of unrelated domains behind their IP addresses, so connecting to their IPs reveals essentially nothing, but if your IP is only used for a handful of things then it's pretty obvious what's going on even if the SNI is obscured.
The popular HTTP validation method has the same drawback whether using DNS or IP certificates? Namely, if you can compromise routes to hijack traffic, you can also hijack the validation requests. Right?
> IP addresses also are assigned by registrars (ARIN in the US and Canada, for instance).
To be pedantic for a moment, ARIN etc. are registries.
The registrar is your ISP, cloud provider etc.
You can get a PI (Provider Independent) allocation for yourself, usually with the assistance of a sponsoring registrar. Which is a nice compromise way of cutting out the middleman without becoming a registrar yourself.
You can also become a registrar yourself - at least, RIPE allows it. However, fees are significantly higher and it's not clear why you'd want to, unless you were actually providing ISP services to customers (in which case it's mandatory - you're not allowed to use a PI allocation for that)
The biggest modern-era reason is direct access to update your RPKI entries.
But this only matters if you are doing stuff that makes direct access worthwhile.
If your setup is mostly "set and forget" then you should just accept the lag associated with needing to open a ticket with your sponsor to update the RPKI.
Very very true, never thought about orgs like that. However, I don't think someone should use this like a bandaid like that. If the idea is that you want to have a domain associated with a service, then organizationally you probably need to have systems in place to make that easier.
Ideally, sure. But in some places you're what you're proposing is like trying to boil the oceans to make a cup of tea
VBA et al succeeded because they enabled workers to move forward on things they would otherwise be blocked on organizationally
Also - not seeing this kind of thing could be considered a gap in your vision. When outsiders accuse SV of living in a high-tech ivory tower, blind to the realities of more common folk, this is the kind of thing they refer to.
One reason is that the client certificate with id-kp-clientAuth EKU and a dNSName SAN doesn't actually authenticate the client's FQDN. To do that you'd have to do something of a return routability check at the app layer where the server connects to the client by resolving its FQDN to check that it's the same client as on the other connection. I'm not sure how seriously to take that complaint, but it's something.
What's stopping you from creating a "localhost.mydomain.com" DNS record that initially resolves to a public IP so you can get a certificate, then copying the certificate locally, then changing the DNS to 127.0.0.1?
Something about a 6 day long IP address based token brings me back to the question of why we are wasting so much time on utterly wrong TOFU authorization?
If you are supposed to have an establishable identity I think there is DNSSEC back to the registrar for a name and (I'm not quite sure what?) back to the AS.for the IP.
Then it would be a grave error to issue an IP cert without active insight into BGP. (Or it doesn't matter which chain you have.. But calling a website from a sampling of locations can't be a more correct answer.)
(seems to be WIP https://github.com/caddyserver/caddy/issues/7399)
This is no criticism, I like what they do, but how am I supposed to do renewals? If something goes wrong, like the pipeline triggering certbot goes wrong, I won't have time to fix this. So I'd be at a two day renewal with a 4 day "debugging" window.
I'm certain there are some who need this, but it's not me. Also the rationale is a bit odd:
> IP address certificates must be short-lived certificates, a decision we made because IP addresses are more transient than domain names, so validating more frequently is important.
Are IP addresses more transient than a domain within a 45 day window? The static IPs you get when you rent a vps, they're not transient.
Nobody is being forced to use 6-day certs though, when the time comes LE will default to 47 days just like everyone else.
Though if I may put on my tinfoil hat for a moment, I wonder if current algorithms for certificate signing have been broken by some government agency or hacker group and now they're able to generate valid certificates.
But I guess if that were true, then shorter cert lives wouldn't save you.
Probably not. For browsers to accept this certificate it has to be logged in a certificate transparency log for anyone to see, and no such certificates have been seen to be logged.
TLS certs should be treated much more akin to SSH host keys in the known hosts file. Browsers should record the cert the first time they see it and then warn me if it changes before it's expiration date, or some time near the expiration date.
Obviously you might still be victim #1 of such a scheme... But in general the CA's now aren't really trusted anymore - the real root of trust is the CT logs.
6 days actually seems like a long time for this situation!
If I don't assign an EIP to my EC2 instance and shut it down, I'm nearly guaranteed to get a different IP when I start it again, even if I start it within seconds of shutdown completing.
It'd be quite a challenge to use this behavior maliciously, though. You'd have to get assigned an IP that someone else was using recently, and the person using that IP would need to have also been using TLS with either an IP address certificate or with certificate verification disabled.
Which should push you to automate the process.
I think acme.sh supports it though.
We also support ACME profiles (required for short lived certs) as of v1.18 which is our oldest currently supported[1] version.
We've got some basic docs[2] available. Profiles are set on a per-issuer basis, so it's easy to have two separate ACME issuers, one issuing longer lived certs and one issuing shorter, allowing for a gradual migration to shorter certs.
[1]: https://cert-manager.io/docs/releases/ [2]: https://cert-manager.io/docs/configuration/acme/#acme-certif...
There would be no way of determining that I can connecting to my-organisation's 10.0.0.1 and not bad-org's 10.0.0.1.
But what risks are attached with such a short refresh?
Is there someone at the top of the certificate chain who can refuse to give out further certificates within the blink of an eye?
If yes, would this mean that within 6 days all affected certificates would expire, like a very big Denial of Service attack?
And after 6 days everybody goes back to using HTTP?
Maybe someone with more knowledge about certificate chains can explain it to me.
With a 30 day cert with renewal 10-15 days in advance that gives you breathing room
Personally I think 3 days is far too short unless you have your automation pulling from two different suppliers.
How many "top of chain" providers is letsencrypt using? Are they a single point of failure in that regard?
I'd imagine that other "top of chain" providers want money for their certificates and that they might have a manual process which is slower than letsencrypt?
But in general, one of the points of ACME is to eliminate dependence on a single provider, and prevent vendor lock-in. ACME clients should ideally support multiple ACME CAs.
For example, Caddy defaults to both LE and ZeroSSL. Users can additionally configure other CAs like Google Trust Services.
This document discusses several failure modes to consider: https://github.com/https-dev/docs/blob/master/acme-ops.md#if...
It depends. If the ACME client is configured to only use Let’s Encrypt, then the answer is yes. But the client could fall-back to Google’s CA, ZeroSSL, etc. And then there is no single point of failure.
To use it, you need a valid certificate for the connection to the server which has a hostname that does get broadcast in readable form. For companies like Cloudflare, Azure, and Google, this isn't really an issue, because they can just use the name of their proxies.
For smaller sites, often not hosting more than one or two domains, there is hardly a non-distinct hostname available.
With IP certificates, the outer TLS connection can just use the IP address in its readable SNI field and encrypt the actual hostname for the real connection. You no longer need to be a third party proxying other people's content for ECH to have a useful effect.
Even if it did work, the privacy value of hiding the SNI is pretty minimal for an IP address that hosts only a couple domains, as there are plenty of databases that let you look up an IP address to determine what domain names point there - e.g. https://bgp.tools/prefix/18.220.0.0/14#dns
> In verifying the client-facing server certificate, the client MUST interpret the public name as a DNS-based reference identity [RFC6125]. Clients that incorporate DNS names and IP addresses into the same syntax (e.g. Section 7.4 of [RFC3986] and [WHATWG-IPV4]) MUST reject names that would be interpreted as IPv4 addresses.
Actually the main benefit is no dependency on DNS (booth direct and root).
IP is a simple primitive, i.e. "is it routable or not ?".
To be pedantic for a moment, ARIN etc. are registries.
The registrar is your ISP, cloud provider etc.
You can get a PI (Provider Independent) allocation for yourself, usually with the assistance of a sponsoring registrar. Which is a nice compromise way of cutting out the middleman without becoming a registrar yourself.
The biggest modern-era reason is direct access to update your RPKI entries.
But this only matters if you are doing stuff that makes direct access worthwhile.
If your setup is mostly "set and forget" then you should just accept the lag associated with needing to open a ticket with your sponsor to update the RPKI.
There's also this little thing called DNS over TLS and DNS over HTTPS that you might have heard of ? ;)
VBA et al succeeded because they enabled workers to move forward on things they would otherwise be blocked on organizationally
Also - not seeing this kind of thing could be considered a gap in your vision. When outsiders accuse SV of living in a high-tech ivory tower, blind to the realities of more common folk, this is the kind of thing they refer to.
* An outer SNI name when doing ECH perhaps
* Being able to host secure http/mail/etc without being beholden to a domain registrar
Other than basically being a pain in the ass.
For local /network/ development, maybe, but you’d probably be doing awkward hairpin natting at your router.
If you are supposed to have an establishable identity I think there is DNSSEC back to the registrar for a name and (I'm not quite sure what?) back to the AS.for the IP.