LetsEncrypt Outage
letsencrypt.status.io
letsencrypt.status.io
Example OpenBSD /etc/acme-client.conf:
authority buypass {
api url "https://api.buypass.com/acme/directory"
account key "/etc/acme/buypass-privkey.pem"
contact "mailto:youremail@example.com"
}
domain example.com {
domain key "/etc/ssl/private/example.com.key"
domain full chain certificate "/etc/ssl/example.com.pem"
sign with buypass
}A pity that acme-client(1) does not allow for fallbacks, but I will add a mental note about it being an easy enough patch to contribute if I ever find the time.
I wrote this blog post a few weeks ago: "Minimal, cron-ready scripts in Bash, Python, Ruby, Node.js (JavaScript), Go, and Powershell to check when your website's SSL certificate expires." https://heiioncall.com/blog/barebone-scripts-to-check-ssl-ce... which may be helpful if you want to roll your own.
(Disclosure: at Heii On-Call we also offer free SSL certificate expiration monitoring, among other things.)
So, what gives?
Which is disappointing because you should be able to recreate the service they had nearly exactly with certificate transparency logs.
Shouldn't that happen automatically a bit beforehand?
https://letsencrypt.org/2024/04/25/guide-to-integrating-ari-...
The team lead couldn't get over the slogan "devops, automating downtimes since 2010" whenever someone wanted to add a new nonessential automation that does things on prod servers.
I mean he wasn't completely wrong, it was a non essential automation with high risk and very little reward (<1h saved every 2 yrs), which is why we never switched to LE for our main site, only internal tooling was allowed to use it
https://www.digicert.com/blog/tls-certificate-lifetimes-will...
I was merely retelling an anecdote about how LE was always positioned to be exclusively about refreshing certs automatically, though. As I've moved out of (dev-)ops roles around 2016/2017 so I'm really not up to date with operations topics
And?
Automation sometimes breaks, both for internal reasons (OS patching) or external. For the latter, LE at some point in the past changed CDNs, and this caused JWST headers to be sent back differently, which broke different clients, e.g.:
* https://community.letsencrypt.org/t/jws-has-no-anti-replay-n...
* https://github.com/dehydrated-io/dehydrated/issues/684
Being able to get e-mails was an extra level of monitoring that was handy, even if you had automation.
And you set up your own monitoring systems for your own infrastructure, as you have always done.
Or better yet, set up auto-renewal as per vendors recommendation.
Vendors - especially vendors you aren't paying - may provide some reminder services, but assuming those to be your sole method for 'managing' your renewals is a deeply poor operational position.
This is going to get really important as cert longevity gets reduced, eg https://news.ycombinator.com/item?id=43693900
If you're using Prometheus or Prom-friendly systems - https://github.com/ribbybibby/ssl_exporter
It can send alerts to multiple alerting providers.
This sounds like an easy problem to identify root cause for.
I think I received about 15 'we're disabling email notifications soon' emails over the past several months - one of which was interesting, but none were needed, as I'd originally set this up, per documentation, to auto-renew every 30 days.
Perhaps create a calendar reminder for the short term?
You should have it on auto-renewal anyway.
You can grab a cert from ZeroSSL and probably some others.
You can also get 1year cert from aws for like $15 though I'd stick with auto renews.
There's no way it's DNS
It was DNS
1. Denial: It’s not DNS.
2. Anger: What the fuck is it!
3. Bargaining: Maybe it’s a firewall, or Cloudflare!
4. Depression: We’ve checked everything…
5. Acceptance: It’s DNS.
(Or, as I recently encountered, it can also be a McAfee corporate firewall trying to be helpful by showing a download progress bar in place of an HTTP SSE stream. I was sure that was being caused by MTU, but alas no.)
As they move to shorter-lifetime certs (6 days now https://letsencrypt.org/2025/01/16/6-day-and-ip-certs/?utm_s...) this puts it in the realm of possibility that an incident could impact long-running services.
That said, I'm unbelievably grateful for the great product (and work!) LetsEncrypt has provided for free. Hope they're able to get their infrastructure back up soon.
Here is the HN announcement: https://news.ycombinator.com/item?id=8624160
Announcement "animated" https://hn.unlurker.com/replay?item=8624160
Subscribers will be able to optin to short-lived certificates via a certificate profile mechanism being added to our ACME API.
We hope to make short-lived certificates generally available by the end of 2025.
The earliest short-lived certificates we issue may not support IP addresses, but we intend to enable IP address support by the time short-lived certificates reach general availability.
https://zerossl.com/ (90 days)
https://www.buypass.com/ (180 days)
One notable exception is Cloudflare: They famously no longer rely solely on LetsEncrypt.
Especially something that needed to be renewed every 90 or is it 40 days now. How about issuing 100 years certificates as a default?
https://cloud.google.com/certificate-manager/docs/public-ca-... (EDIT: Google is their own CA, with https://pki.goog/ )
The browsers and security people have been pushing towards shorter certs, not longer ones. Knowing how to rotate a cert every year, if not shorter, helps when your certificate or any of your parent certs are compromised and require an emergency rotation.
In a practical sense you likely wouldn't like the alternatives, because for most people's usage of the internet there's exactly one authority which matters: the local government, and it's legal system - i.e. most of my necessary use of TLS is for ecommerce. Which means the ultimate authority is "are you a trusted business entity in the local jurisdiction?"
Very few people would have any reason to ever expand the definition beyond this, and less would have the knowledge to do so safely even if we provided the interfaces - i.e. no one knows what safety numbers in Signal mean, if I can even get them to use Signal.
Your users that visit your website and get a TLS warning are the authority to worry about, if you're running a business that needs security. Depending on what you're selling, that one user could be a gigantic chunk of your business. Showing your local government that you have a process in place to renew your TLS certificates, and your provider was down is most likely going to be more than enough to indemnify you for any kind of maliciousness or ignorance (ignorantia juris non excusat). Obviously, different countries/locations have varying laws, but I highly doubt you'd be held liable for such a major outage for a company that is in such heavy use. Honestly, if you were held liable, or think you would be for this type of event, I'd think twice about operating from that location.
Caddy uses ZeroSSL as a fallback if Let’s Encrypt fails!
I'm using Caddy here and it's not falling back on ZeroSSL. Thanks for your help
EDIT: hmm, it should be automatic...! https://caddyserver.com/docs/automatic-https#issuer-fallback interesting, I'll double check my config
woah... it's probably related to this! https://github.com/caddyserver/caddy/issues/7084 TLDR: "Caddy doesn't fall back to ZeroSSL for domains added using API" (which is my case)
https://github.com/lucaslorentz/caddy-docker-proxy/blob/mast...
caddy-docker-proxy generates a Caddyfile in memory, and then it calls the internal Caddy API `/load`. So - yes, it would presumably have been affected, at least from what I saw yesterday (with our own Caddy) and if the github issue I mentioned is correct.
Also, you have days to weeks of slack time for renewals. The only real impact is trying to issue new certs if you are solely dependent on LE.
The whole point of the expiration is in case a hacker gets the private key to the cert and can then MITM, they can keep MITMing successfully until the cert the hacker gives to the clients expires (or was revoked by something like OCSP, assuming the client verifies OCSP). A very long expiration is very bad because it means the hacker could keep MITMing for years.
The way things like this work with modern security is ephemeral security tokens. Your program starts and it requests a security token, and it refreshes the token over X time (within 24 hrs). If a hacker gets the token, they can attack using it until 1) you notice and revoke the existing tokens AND sessions, or 2) the token expires (and we assume they for some reason don't have an advanced persistent threat in place).
Nobody puts any emphasis on the fact that 1) you have to NOTICE THE ATTACK AND REVOKE SHIT for any of these expirations to have any impact on security whatsoever, and 2) if they got the private key once, they can probably get it again after it expires, UNLESS YOU NOTICE AND PLUG THE HOLE. If you have nothing in place to notice a hacker has your private key, and if revocation isn't effective, the impact is exactly the same whether expiration is 1 second or 1 year.
How many people are running security scans on their whole stack every day? How many are patching security holes within a week? How many have advanced software designed to find rootkits and other exploits? Or any other measure to detect active attacks? My guess is maybe 0.0001% of you do. So you will never know when they gain access to your certs, so the fast expiration is mostly pointless.
We should be completely reinventing the whole protocol to be a token-based authorization service, because that's where it's headed. And we should be focusing more on mitigating active exploits rather than just hoping nobody ever exploits anything. But that would scare people, or require additional work. So instead we let like 3 companies slowly do whatever they want with the entire web in an uncoordinated way. And because we let them do whatever they want with the web, they keep introducing more failure modes and things get shittier. We are enabling the enshittification happening in front of our eyes.
In old-school X.509 PKI this might be "in case this person is no longer affiliated with the issuer" (for organizational PKI) or "in case this contact information for this person is otherwise no longer accurate".
In web PKI this might be "in case this person no longer controls this domain name" or "in case this person no longer controls this IP address".
The key-compromise issue you mention was more urgent for the web PKI before TLS routinely used ciphersuites providing forward secrecy. In that case, a private key compromise would allow the attacker to passively decrypt all TLS sessions during the lifetime of that private key. With more modern ciphersuites, a private key compromise allows the attacker to actively impersonate an endpoint for future sessions during the lifetime of that private key. This is comparatively much less catastrophic.
In terms of "no longer controls this domain name", or "no longer controls this IP address", there are a raft of other issues related to this that expiration doesn't cover:
- Does the real domain owner still have a DNS record pointing to an IP address they no longer own? If yes, attacker that now has that IP can serve valid TLS.
- Does the attacker control either the registrar account, or the name server account, or can poison DNS, or an HTTP server, or an email server, or BGP? If yes, the attacker can make new certs.
There's so many holes in TLS it's swiss cheese. Expiration as security is like a cardboard box as a bulletproof vest. Yet that cardboard box is so bulky and cumbersome it makes normal life worse.
Nope. So all that happened here is that you were wrong.
Note that you can make your own self-signed CA certificate, create any server and client certificates you want signed with that CA cert, and deploy them whenever and wherever you want. Of course you want the root CA private key securely put somewhere and all that stuff.
The only reason it won't work at large without a bit of friction is because your CA cert isn't in the default trusted root store of major browsers (phone and PC). It's easy enough to add it - it does pop up warnings and such on Windows, Android, iOS and hopefully Mac OS X, but they're necessary here.
No, it's not going to let the whole world do TLS with you warning-free without doing some sort of work, but for small scales (the type that Let's Encrypt is often used for anyway) it's fine.
If only the same zest applied to probes
But DNS does not have a good reliable authenticated transport mechanism. I wonder if there was a way to build this that would have worked.
Half the year I live on an island that is reliant on submarine cables and has historically had weeks and months long outages and with a changing world I suspect that might become reality once again. Locally this wasn't much of an issue, the ccTLD continues to function, most services (but now about 35%) are locally hosted. Then HTTPS comes along. Zero certificates could be (re-)issued during an outage. A locally run CA isn't really an option (standalone simply isn't feasible and getting into root stores takes time and money), so you are left with teaching users to ignore certificate errors a few weeks into an extended outage.
I could see someone like LE working with TLD registrars to enable local issuance (with delegated/sub-CA certificates restricted to the TLD), that could also mitigate problems like today (decentralize issuance) and the registrars are already the primary source of truth for DV validation.
Your registrar should be able to validate your ownership of the domain, ergo your registrar should be your CA. Instead of a bunch of arbitrary and capricious rules to be trusted, a CA should not be "trusted" by the browser, but only able to sign certificates for domains registered to it.
Not that the other parties are that independent anyways: Microsoft's browser is a Google fork, and is wholly dependent on it. Mozilla's entire funding is Google. Apple is arguably the only somewhat independent party here, but that multibillion dollar annual search deal... let's say it incentivizes collaboration.
And the push-down in certificate age is, or at least was at the beginning, a push from Apple. The others have come around/along for the ride
The "I support shorter lifetimes so this all comes crashing down" comment I made earlier is arguably a bit facetious, but I do think the PKI wonks in the CAB are pretty much accountable to noone until they break things badly enough that their bosses have to pay attention to the problem.
Antitrust enforcement remains the fix here.
Ultimately the problem is that currently "security best practice" as it's commonly discussed, says what we're doing now is a good idea. It's not, and until we change the understanding on that, nobody's going to feel motivated to do better.
Password rotation used to be considered a gold standard strategy for security, until people realized not only did it make everything harder, it also encouraged people to choose less secure passwords and was largely self-defeating.
If I told you we could improve a 90-day password rotation policy by making it change every week, you'd rightly call me crazy, but for some inconceivable reason (monopoly, perverse incentives, appeal to an authority run by idiots, name your choice), people act like decreasing certificate lifetime is somehow going to make the web safer.
But whether that's intuitive for you or not, the fact remains: short-lifetime automated certificate provisioning is a response to revocation.
I think short certificate lifetimes will be viewed in the relatively short future as misguided as the old NIST recommendation on passwords.
To understand the distinction, imagine a password system where you can't change your password. You can make new ones, but the old ones still work. That's the problem TLS was facing.
PKI is structurally broken at every level. The reason there are so many changes and additions and new features being slapped on it, is because it's a really dumb system and the people who control it are very motivated to keep it in place at the expense of everything else.
Even if we ignore the fact that certificates are not a secret, and that expiry applies to certificates, not private keys, a major difference is that humans don’t mentally generate or manually type TLS keys or certificates. So the negative impact of rotation on user experience and behavior is entirely absent.
Also, short lifetime certificates help deprogram concern about certificate warnings (most nontechnical users know to ignore them, as a network admin, I've never seen a certificate warning that was actually due to a compromise... so I also ignore them all), which leads to hypothetically much less safe behavior than if certificate warnings only happened when rational.
Which is to say, if you believe a certificate that expired yesterday should result in a scare screen to users or worse with HSTS, interfering with the ability to access it all, you failed security 101.