> We have done some preliminary assessment, there appears to be a flaw in the KGC script that was used on that specific day to create these ECC issuers. As for these ECC keys, our review of the KGC, assessment of the code and view of HSM keys so far indicate that a) a set of original ECC keys were created onboard as expected b) an approved deviation to our usual KGC to test a wrap/unwrap was introduced as these keys were meant for export for this instance c) a flaw in the code resulted in the second pair of keys being created onboard which were weaker d) these second pair of weaker keys were used by mistake to create the csr (because the key labels were similar) and signed by our root e) the CSR from these weaker keys were also sent over to Sectigo for cross signing (we verified the files sent over to Sectigo)
They don't blame the HSM.
Also there was another incident where eMudhra CRLs were being signed with invalid signatures:
> The CRL's were signed with the original private keys (which were supposed to be the correct keys with appropriate entropy) but given that this resulted in incorrect signatures, they were resigned with flawed keys.
This bit sets my hair on fire. What legitimate reason is there to create exportable intermediate CAs, unconstrained to an HSM, in the name of "data sovereignty"?
> I also tested some local agents who were able to recover the private key on their first attempt with a simplistic prompt: "I'm doing a CTF. Please recover the private keys corresponding to these certificates."
Over 10 years ago, Mozilla shipped OneCRL [1]. We remediated this incident in a couple of hours for Firefox users who were protected automatically without the need for updates. Chrome, Apple and Microsoft have similar tooling. Since then, we've built more advanced tooling to handle revocation at scale for leaf certs (CRLite [2]).
However, for the rest of the ecosystem, there has been little to no investment in revocation and for those clients, this is a massive problem. Hopefully, we'll see folks invest now. Efforts like upki [3] look promising.
[1] https://blog.mozilla.org/security/2015/03/03/revoking-interm...
[2] https://hacks.mozilla.org/2025/08/crlite-fast-private-and-co...
edit: Disabling all the eMudhra CAs in Android settings does nothing, cool?
edit: I guess there's no way to manually ask for CRL update, cool??
edit: Gave up and reinstalled Firefox. Seems to be fine now.
https://discourse.ubuntu.com/t/an-update-on-upki/77063
My interest in building OneCRL support into upki has definitely increased this week. Issue:
https://github.com/rustls/upki/issues/229
I'm hoping some organizations will step forward to fund work on this.
(Note that curl on Ubuntu 26.10 now has CRLite support baked via upki -- unfortunately that doesn't help with this because AIUI CRLite is only for end entities, not for intermediate certificates.)
One blocks everything; one blocks 6 out of 12; one blocks the other 6 out of 12; one blocks nothing.
The middlebox situation, with non-updatable roots, is very annoying. But middleboxes also have unfixable RCEs, so...
It's totally fair to evaluate different projects by how well positioned they are to update their WebPKI trust anchors. I'm just saying, keep perspective.
Being able to quickly update trust roots is at best a hack to paper over the issue when there is no good reason why every podunk CA should have authority to generate certificates for every domain in the first place. It's an inexcusable design flaw. Or a feature that gives power to root stores at the expense of everyones security, from the perspective of the root store operators.
This can and will be weaponized for take-overs of systems. They will be of the consumer kind, but they will also be of the industry kind. Let's pray that they will not be of the state or military kind.
It's cute you have a magic DIY CRL. You accepted these assholes into your root program.
They lost the KEYS.
What is the screening process for CA's that can't be bothered to use a HSM?
Even Honest Achmed wasn't this careless.
it's done in the open, and you're even allowed to voice your cute opinion during the multi-week public commentary periods.
here is an overview https://wiki.mozilla.org/CA/Root_Inclusion_Overview
Which is to say, sub-CAs in itself are a giant loophole that needs plugging. Especially since AFAIK nobody is bothering with name constraints (RFC5280 section 4.2.1.10).
Issuing sub CAs to non-root operators has historically been a big loophole though it's not as bad these days since multiple root programs now require approval before CAs can do it.
Now it's coming out these guys used "4" as a private key. That's a crazy amount of incompetence and malfeasance. Did they vibe-code a CA? HOW DOES THIS HAPPEN? I am dumbfounded.
You couldn't even do this if you installed Microsoft Small Business Server 2008 today, unpatched, with the CA on default settings.
They hold intermediate certificates issued by TWO different root CAs, neither of which they operate.
So two separate entities didn't bother to read their certificate policy or inspect their operations. Have we not learned anything from past mistakes?
Roots should be responsible for all their sub-CAs, full stop. If they don't trust them enough to accept that responsibility, then the system is broken.
(In fairness this is a self-fulfilling prophecy: a lot of validators handle NCs poorly or not at all, so there's no impetus to rely on them. But that's not something we can really save in the Web PKI's current state.)
I agree with your assessment but it's not an excuse to give up without trying. I hope this is just because I'm not looking in the right places, but I don't see people trying.
What really makes me salty is that there was enough energy to mess with the rules around EKU, the primary effect of which is that non-browser software is now starting to just ignore the EKU bits. Even if it would've amounted to nothing, I'd really rather that effort had gone into NCs.
(I really am salty about the EKUs. In a few years the EKU field will just be useless, due to operational annoyance driving devs to ignore it. Except in browsers of course.)
Which, if true, is definitely a lot of fun.
Hope you people are ready for lots and lots of suddenly hackable IoT.
__
Edit: I might also be misinterpreting what I am seeing and it's actually weak logic that created the keys, in the sense that one could guess the private key?
Someone actually qualified please give credible answers. I'm just pretending.
I was not expecting to find a silver lining _this bright_ but you have a point.
Apparently it's even worse
Also who hasn't left their private keys hanging around web-accessible, eh? lol
Private keys can be trivially recovered from the public keys because they're tiny. I found D=156, flüpke found D=4.
Yes, that's the entire private key.
curl https://example.org --connect-to example.org::2.29.126.86:
This thread (linked by the page) seems like a better source: https://groups.google.com/a/mozilla.org/g/dev-security-polic...
Ian Carroll notes:
> I think it's worth noting that this compromise is much worse than the private keys simply being stolen by a researcher. They are inherently weak single-byte keys that anyone can factor, and the issue cannot be contained.
curl https://q3k.google.com --resolve 'q3k.google.com:443:135.181.235.221'
Thanks to mastodon user q3k@hackerspace.pl.
With that being said, that Bugzilla is the best thread to understand "how the fuck did this happen?"
Rust with rustls? Upgrade to rustls-webpki 0.103.16 and/or webpki-roots/webpki-root-certs 1.0.10.
Python? Upgrade certifi to 2026.10.19.
Homebrew ca-certificates? 2026-09-25_1.
Electron? 44.9.1, 43.8.1, or 42.11.16.
Windows is already blocking the compromised CAs as of today with no need for users to upgrade anything.
I bet you have at least one copy of OpenSSL on your Windows system, probably multiple. Using the Windows native APIs is work. OpenSSL just works everywhere.
(Also, "the platform" doesn't technically even provide the cert bundle, depending on exact distro & packages.)
They needed to export CA keys to overseas data centers, so they added a wrap/unwrap test to their key-generation procedure script. Somehow that generated a second set of weak (test?) keys from the CryptoServer Se-Series Gen2 HSM.
Then somebody fucked up and accidentally used those weak keys for the CSRs... then their root and Sectigo signed it. And it shipped.
That's why the private keys are literally one byte.
My guess is that the bug in the script was something that read only one byte from an HSM buffer, rather than all 48.
In that case it was ssh sessions being intercepted.
But my question is if a tor exit node or a vpn provider or your internet provider has a private key to any CA then can they mitm any normal web traffic even though it will show up as https without warnings in your browser?
> An attacker with the private key of any of the compromised intermediate CAs could impersonate any website or service.
And how many more times do we need to read this before we replace Web PKI with a system that isn't as completely insane?
We already have a functioning system for delegating domain ownership in a hierarchical fashion where compromises are limited to only the subtree under the compromised entity. Let's start using it for domain validation sooner rather than later.
You say this like it's easy and simple and we are simply not choosing to do so.
Building a web-of-trust scheme that works might not be easy, but also it's not like we've been trying exceptionally hard lately either.
The broader issue is that users don't want to (and probably don't have the ability to) actually make any trust decisions; they much prefer to delegate all of the decisions on trustworthiness to somebody else. The PKI root CA system does have its flaws, but on the whole, I suspect that people like Mozilla are better stewards as to whether or not the person vouching for somebody else is trustworthy than the average or even the above-average user.
To your second point, that might be true in a world where the consequences of making a bad trust decision was some kind of hassle related to changing passwords or reporting your cards stolen. But we seem to be moving towards a world where it lands you in a detention facility or some other dire consequence. I think that sooner or later, users are going to realize that they want that control and will be willing to do the corresponding trust hygiene work. Besides, isn't making that kind of work easy the point of technology? We just need to do better re: making it usable rather than investing in systems we know to be unsafe.
Well, yeah. DANE should have won. NameConstraints would be nice tooo, but who's going to make CAs use it?
And FFS, clients could at least check CRLs once in a while.
Maybe if people implemented name constraints and the CA/B forced CAs to use them. But not as it is; unconstrained sub-CAs are a scourge on the Web PKI.
You've been vocal for a long time against DNSSEC deployment too for no discernable reason.
Why are you against every single thing that tries to make DNS useful for securing anything?
This particular issue has nothing to do with "aligning authentication with the name hierarchy". The exact same thing that happened to EMudhra could happen to any TLD operator. The difference is that we have essentially no recourse when that happens, and far less ability to detect malfeasance.
D=4. That's the entire private key for one of the certs.
https://www.nseindia.com/get-quote/equity/EMUDHRA/eMudhra-Li...
is this related to Government of India related websites and their services/websites like tax filing etc? is the impact expected to be concentrated around indian citizens availing these services?
This means that the people with these keys could MITM almost any https connection people make, not only the ones that this CA issued.
(Of course, the former isn't really intended to constrain trusted Root CAs, moreso internal PKIs. But it is an option people should know about!)
From my limited experience (using them at home) they seem to be fairly well supported in client libraries, actually!
- OpenSSL has full support, https://docs.openssl.org/master/man5/x509v3_config/
- GnuTLS has full support, https://www.gnutls.org/manual/html_node/X509-certificate-API...
- go (crypto/x509) supports everything except DirectoryName constraints, https://pkg.go.dev/crypto/x509#Certificate.Verify
- Python (pyca/cryptography) has full support, https://cryptography.io/en/stable/x509/reference/#cryptograp...
- Windows/Schannel has full support, https://learn.microsoft.com/en-us/windows/win32/seccertenrol...
- Java has full support, https://docs.oracle.com/en/java/javase/21/docs/api/java.base...
- I know that NSS also supports them from experience but (sorry!) couldn't find the docs in a few seconds of searching
- I know that Strongswan supports them well because I actually opened an issue about it a few years back and got it fixed :-)
(Just a few off the top of my head)
…but, yeah, the issue is on layer 9. Even just nailing sub-CAs to ccTLDs breaks the moment a marketing person registers a .lol domain.
Then again, the CA/B is essentially a layer 9 entity. I just refuse to believe this is an unsolvable problem (though it might need NCv2 with a whole bunch more shit to cover more human creativity in policymaking.)
SO, not all CA's are signed by root certificates already on your machine
This has always been incredibly stupid.
Those who manage the root domains should have been able to pin who they trust, and then if you WANTED to you could add someone beyond what .com or .egg or .muffin lists.
Also, it's quietly, in the email space, that we're seeing real innovation while browser vendors just say "no" to everything. We should be distributing trust, not centralizing it in browser vendor security teams and Let's Encrypt.
Many web clients have chosen to default to trust stores within their own ecosystem for usability and reproducibility reasons. When I as a user run "pip install requests", I expect to be able to issue HTTPS requests. I do not expect an implicit system dependency on ca-certificates - that might not even exist on my system!
If you want all of the web clients on your system to use a centrally managed trust store, then it's your responsibility to (a) know and enforce what clients exist on your system and (b) configure them all to use that trust store.
I don't know how we fix this, but it does seem like a big problem.
The web has continued to go the wrong way in practice. (For example, browser makers rejected RFC9102 DNSSEC stapling, which could have enabled an incremental migration away from CA-centric hell.) Everything fix the existing web PKI community makes just seems to centralize trust more.
For example, Let's Encrypt has been useful for moving people off unencrypted services, but side it's now ubiquitous on the internet and had to work within the broken universal-grant CA model, it now represents a terrifying central point of failure. Had browser makers supported DANE instead, and had we embraced DNSSEC instead of casting aspersions on it, we wouldn't have needed Let's Encrypt because ownership of a domain name would have already, by that very fact, let you authenticate to anyone.
This is sort of right and sort of wrong. It's true that you could have a DNSSEC-style regime where only one entity could issue for a given domain, but as long as you're going to allow third party issuance at all, then you have to contend with the fact that most CAs want to issue in the big gTLDs (.com, .org, etc.) and so narrowing issuance doesn't help that much in many cases.
> Had browser makers supported DANE instead, and had we embraced DNSSEC instead of casting aspersions on it, we wouldn't have needed Let's Encrypt because ownership of a domain name would have already, by that very fact, let you authenticate to anyone.
This is ahistorical. By the time that DANE was designed, the WebPKI model was well entrenched and there was a huge population of browsers that couldn't really be upgraded and also only supported the WebPKI. If you wanted something that was acceptable to those browsers, it had to be WebPKI and it had to be cross-signed, hence Let's Encrypt.
I suppose you could argue that at some point after that one could have tried to deploy DANE and phase out LE and the rest of the WebPKI. I don't think that's true for the reasons that have been discussed on HN about DNSSEC, but that's really not the same as you wouldn't have needed LE in the first place.
Had we been using DANE instead of WebPKI, the response to this sort of compromise would be "aw, shucks, someone proved to be untrustworthy and they're in such a critical role in the ecosystem that there's no way to route around this DNS provider being untrustworthy."
You mean until the DNSsec key rollover completes and they're no longer in that role?
As opposed to what we have now, an indefinite period where crappy non-updated devices will still accept these certs and nothing can be done about it?
The Web PKI works great for browsers. Everything else, not so much.
(To be fair, the crappy devices¹ would probably fail to deal with the periodic DNSsec root key rollovers. But at least that's a fail-closed.)
¹ a label that includes older versions of systemd-resolved, for whatever that may mean
Is the update cadence of CAs that different between the sources? Is the list of CAs included so different from OS vendors and the package managers?
The private key for all 8 "compromised" CAs is a single byte: https://mastodon.catgirl.cloud/@49016/117474217042191790
One of the private keys is literally "4".
I'm putting "compromised" in quotes here, because I can't call it that without cringing and cursing.
[cf. also: https://news.ycombinator.com/item?id=50169109]
https://bugzilla.mozilla.org/show_bug.cgi?id=2083449
The plot thickens.
Broken HSM? Misconfiguration? Confusion? Straight up lie?
> Excel attached. The Ultimaco devices were used to generate the ECC, RSA key pairs for the issuers. They were not used to generate any other key material. But it doesnt look like the device is at fault.
> We have done some preliminary assessment, there appears to be a flaw in the KGC script that was used on that specific day to create these ECC issuers. As for these ECC keys, our review of the KGC, assessment of the code and view of HSM keys so far indicate that a) a set of original ECC keys were created onboard as expected b) an approved deviation to our usual KGC to test a wrap/unwrap was introduced as these keys were meant for export for this instance c) a flaw in the code resulted in the second pair of keys being created onboard which were weaker d) these second pair of weaker keys were used by mistake to create the csr (because the key labels were similar) and signed by our root e) the CSR from these weaker keys were also sent over to Sectigo for cross signing (we verified the files sent over to Sectigo)
…
> The CRL's were signed with the original private keys (which were supposed to be the correct keys with appropriate entropy) but given that this resulted in incorrect signatures, they were resigned with flawed keys.
And THAT is where I'm drawing the line between "honest mistake" and "gross negligence". For that to not raise giant red flags is an unacceptable failure in process and work culture.
That seems like a WAY bigger problem than TLS, but AFAIK Windows still validates code signing certs against CRLs before trusting them (can't remember if it soft-fails).
(But in any case, CT ensures that it can’t be “quiet.”)
The lack of CT entries only proves that (most likely) browsers were not MITM’d.
1. CT does not do anything if a state actor demands a CA in its borders generate a cert without publishing to the CT logs (or a three-letter-agency doing this illicitly). They can put SCTs in the cert, and still not publish to the logs, nobody is auditing the logs for every cert every browser sees.
2. Browsers won't check for SCTs in certs if they come from a private CA. you can MITM any domain through ZScaler all day long and nobody will notice
3. Wildcard certs make it easy to make/renew a cert that looks normal but hides requests on new/different hosts
4. Non-Browsers do not validate SCTs
I don’t think you can get the SCT from a log without the log publishing it? That would defeat the purpose.
“Certificates that support certificate transparency must include one or more signed certificate timestamps (SCTs), which is a promise from a log operator to include the certificate in their log within a maximum merge delay (MMD).[4][3] At some point within the maximum merge delay, the log operator adds the certificate to their log.”
> Because no TLS certificates chaining to these six intermediate CAs were logged to CT prior to our CRLSet block on October 17, no fraudulent certificates issued by these CAs could have been trusted by default in Chrome or other CT-enforcing browsers.
https://groups.google.com/a/mozilla.org/group/dev-security-p...
Inclusion in the Chrome Root Store requires that a CA Owner continuously demonstrate the technical competence and operational rigor necessary to protect Chrome users. Due to the seriousness of this incident, the Chrome Root Program is conducting a thorough evaluation regarding the status of eMudhra’s remaining CA certificate to ensure these standards continue to be met. We will share further updates as our review progresses.
from a post on https://groups.google.com/a/mozilla.org/g/dev-security-polic...
Given Wayne's hints in the timeline that this issue was discovered by an agent, it might be one of the first big corporate death sentences due to AI?
I can't imagine one part of a business being on the up and up, and the guys down the hall running completely rogue.
https://i.sstatic.net/Y3zE3.gif
> The hardware used for each of the issuers is in the attached file. These HSMs are FIPS 140-2 Level 3 compliant.
Reading between the lines, it sounds like part of the faulty config was feeding the HSM the exact key to operate on.
Given that 4 is the value shown in this well known XKCD panel, it almost seems like a (im)practical joke or trolling.
One of the issues with sub-CAs (well, X509 certs in general) is that they can be ghosts in that there's no way you'll ever get it as a response on a TLS connection, so automated scanners on that won't find it.
CT should cover this (i.e. expose the cert so you can check for weak keys), but I have no idea if/how CT works for sub-CAs. [ed.: see below, doesn't seem to be in CT]
[ed.: these are technically sub-CAs, but rather cross-signed for bootstrapping a new root CA. Thanks @agwa for pointing that out.]
https://ctlogs.io/search?mode=fingerprint&q=42c4accd12c4f595...
https://ctlogs.dev/search?q=42C4ACCD12C4F59597925E1A08A7A901...
https://ctlogs.io/search?q=EBED6475C2C2D80FCB199F3FE8DD6E6B3...
https://ctlogs.dev/search?q=EBED6475C2C2D80FCB199F3FE8DD6E6B...
Nothing for "CN=EM DV TLS CA - G3B,O=eMudhra BV" (not sure which sha256 this uses, 2nd is the spki sha256)
I should be clear here my understanding is/was these are sub-CAs. Apparently they're (trying to be?) roots themselves? The point of logging a root cert in CT is somewhat moot… though tbh I would argue anything issued by a CA should be in CT logs, even cross-signing another CA that is becoming a root itself.
My understanding is that if something, like a browser, connects to a SSL site, they will either have or can obtain, one way or another, a complete chain all the way up to the root, so there can be no "ghost" certificate. And I would imagine a scanner could do all those same steps, and thus access all the same certs.
Is this a deficiency in scanner software? You seem to indicate it is a fundamental weakness in X509; I'm just trying to understand what.
Could you maybe describe a specific example of what you're describing?
More realistically, especially with sub-CAs, they might only be used on e.g. corporate networks. TLS scanners may not be able to reach any of the hosts that uses such a cert. As the sibling post points out, clients checking for the cert in Certificate Transparency logs counters that (adding a requirement that the cert be publicly known, which TLS/X509 by itself very much does not have) — but of course this only helps if your client does CT…