Oct 20, 2026, 3:20 PMOn the eMudhra CA certificate compromise: I just saw they advertise how they store their intermediate certs in HSMs on their website. There is a PoC of signatures from these keys making the rounds, which illustrates pretty well what the limits of HSMs are: To a network attacker, a HSM is little more than an API boundary, and if they can get access to something with access to the HSM, they don't need the keys, they can just tell the HSM to sign what they need to be signed.
Oct 20, 2026, 3:23 PMThere's an interesting detail in that screenshot that's hard to notice to a layperson: They say their HSMs are FIPS 140-2 level 3. FIPS 140-2 has four levels, and what most don't know is that only at level 4, the highest level, is any serious active tamper sensing required. At level 3, which they use here, the requirement is literally:
> Hard opaque potting material encapsulation of [the circuit board] or strong enclosure with removal/penetration attempts causing serious damage.
Oct 20, 2026, 3:24 PMThis means that against a physical attacker, FIPS 140-2 level 3 literally only means "we glued the server shut really well", and all that's needed to get inside is a drill or an angle grinder.
Oct 20, 2026, 4:25 PMhah, it wasn’t a compromise of their infra, it’s way, way worse: apparently somehow, SOMEHOW those intermediates’ private keys had like eight bytes of entropy…between the eight of them, total 🤯
Oct 20, 2026, 4:26 PM(1) how do you fuck up this bad as a CA
(2) what exactly were their auditors doing when they signed off on their baseline requirements compliance
Oct 20, 2026, 4:25 PM@jaseg 8 bits. The key I used had D=156 for instance. :)
@jaseg Assuming the keys might have been generated on a HSM: Is it know which vendor one should extra avoid as a result?
@jaseg On the positive side. Yay transparency systems.
@jaseg That forged cert thus only works with non-CT clients. Which obviously still is a problem because that basically is everything that is not a browser.
But with MTCs that issue will also go away
@jaseg Also I think you could do something like https://git.glasklar.is/sigsum/apps/sign-if-logged for a "proper" HSM too. Probably not using Sigsum but just a generic log of payloads and HSM would only sign things that were sufficiently witnessed before. (My main gripe with the tkey is that there is no attestation path for knowing the provenance of key material it has generated)
Obviously depends on your application of this is desirable of course.
@jaseg Seems like it was broken key generation after all. I wonder if the HSM is broken or if they imported broken keys
Oct 20, 2026, 4:12 PM@23n27 oof! If the HSM was broken that would be a very big deal, because there is not that many manufacturers of these things so that would likely affect a lot of other places as well.
Oct 20, 2026, 4:12 PM@23n27 do you have a link on this being bad keys?