In October 2026 an independent researcher reported that the private keys of eight intermediate certificate authority (CA) certificates operated by eMudhra could be recovered almost instantly. According to eMudhra’s own incident report, the cause was not a weakness in a hardware security module (HSM). It was a custom script written to move CA keys between HSMs from different vendors. This article separates what the primary sources establish from what we, as PKI and HSM consultants, take from them.

How to read this article. The section “Established facts” restates only what named primary sources say, with numbered references. The section “SafeCipher analysis” is our engineering opinion and is labelled as such. We do not allege anything the record does not state. The incident is still under review, and this article reflects the sources as at 16 November 2026.

Established facts

What the incident report says happened

eMudhra published its full incident report on Mozilla’s Bugzilla on 10 November 2026.1 It states that on 30 April 2025, in a single key-generation ceremony, sixteen subordinate CA key pairs were generated on an HSM: eight ECDSA P-384 and eight RSA-4096. The purpose was to move keys between HSMs from different vendors in different locations. The report says this was the first and only time the process had been used.

The report gives the context. The HSMs ordered for the destination facility were delayed by supply-chain constraints. A root-store policy change was due on 15 June 2025, and a cross-certification had to be completed before then. eMudhra therefore generated the keys on the HSMs available and intended to transfer them by key wrapping.

A custom Python script performed the wrap and unwrap. According to the report, after the wrapped key was returned the script read the unwrapped private value out of the module into host memory and created a new key object from that value. For the ECDSA keys, one line in the read-back truncated the private value to its leading byte, leaving a private scalar below 256. That re-created object, not the original, was used to produce the certificate signing request and was certified. The original and re-created objects had almost identical labels. The report states that the HSMs functioned correctly and were not at fault, and that the source module was operated with FIPS-approved mode disabled for the wrap and unwrap step as an approved deviation.1

Mozilla’s analysis describes why this is severe. A P-384 private key should be a random integer between 1 and the group order. These eight were integers below 256, so they can be recovered near instantly by baby-step giant-step, or by generating the first 256 candidate keys.2

Detection and response

  • 17 October 2026: a researcher reported the keys to browser root programmes. Mozilla confirmed proof of key control, recovered the keys independently within hours, and completed a OneCRL block for Firefox by 22:59 UTC.23
  • 18 October: eMudhra revoked the eight ECDSA intermediates. The report says they had issued no end-entity certificates, and that the certified keys had been publicly visible for about 18 months.1
  • 19 October: Sectigo revoked its cross-certificates of those keys.2
  • 20 October: Mozilla removed trust in the root that signed the affected intermediates from NSS to protect software that does not process OneCRL, and released Firefox 158.0.1 and ESR updates.34 Chrome states it blocked the six TLS-issuing intermediates by public-key hash through CRLSet.4
  • 23 October: Utimaco, whose HSMs the report names as the source module, published a statement that, on the information disclosed, a scripting error appears to have created alternative keys used in place of the originals, and that it had found no evidence of a defect in its random-number generation or on-board ECC key generation. This is a vendor statement, not an independent finding.5
  • 30 October: eMudhra appointed KPMG as independent external reviewer.6

Chrome’s Root Program also stated that, before this incident, standard linting tools checked RSA and DSA keys and used blocklists for ECDSA, but did not include baby-step giant-step checks for small elliptic-curve private scalars. It says it ran such checks, using Google’s open-source paranoid_crypto, across Chrome-trusted CA certificates after the report.7

The RSA keys

The report says the eight RSA-4096 keys were not truncated and their signatures verify. However, because they were re-imported from the memory of a Python process on the ceremony host, eMudhra says it cannot positively demonstrate compliance with its own certification practice statement (CPS) section 6.2.6. It treated them as non-compliant and revoked those eight issuers on 7 November 2026. The report states that the four RSA issuers in production (the “G2C” series) had issued end-entity certificates from May 2026, and that eMudhra has decided to exit public TLS issuance.1

How the explanation developed

eMudhra’s first detailed technical comment, on 19 October, said a code flaw had caused “the second pair of keys being created onboard”, that the weaker pair was used “by mistake” because the labels were similar, and that the RSA issuers “went through a different path” and were not impacted.8 The full report of 10 November instead says the faulty object was re-created from a value read out to host memory, and that the RSA keys followed the same read-and-re-create pattern.1 The later, fuller account is the one we rely on. Early root-cause statements in any incident are provisional.

An earlier signal

In May and June 2025 the certificate revocation lists (CRLs) for the G3B and G3C TLS intermediates failed signature verification.9 The report says the CRLs were signed with the originally generated objects while the certificates had been issued over the re-created ones, and that the problem was fixed “at the symptom level” by re-signing with the matching object. eMudhra now describes it as an early indication of the underlying defect.1

What is still open

Independent review by KPMG is in progress, and the report’s own action items run to 4 December 2026. Participants on the public bug have asked for the script, ceremony logs and further detail; the bug remains open and these questions are not answered in the record we reviewed.110 We therefore draw no conclusions about matters the report does not address, including how the FIPS-mode deviation was approved and assessed.

SafeCipher analysis

The points below are our engineering view. They are not statements of what eMudhra should have done, and not statements of what any standard requires.

1. Treat cross-vendor key migration as a design problem, not a script. Moving a CA key between HSMs from different vendors is constrained by wrapping formats, key attributes and export policies. If native mechanisms do not fit, that is the moment to redesign. A custom bridge is where unreviewed code first touches key material. The report itself records, under lessons learned, that the migration created a need for a custom script outside the vendors’ native mechanisms.1

2. Keep private key values inside the HSM boundary. The central failures described involve key material read into host memory. If a design needs a private value to pass through a general-purpose computer, the objective that the key is never exposed outside a validated module has already been lost. Prefer vendor-supported wrapped-key transfer. When a migration looks impractical, generating a fresh subordinate CA key in the destination HSM is a design option, with its own cost: a new CA certificate to issue, cross-sign where needed, and disclose.

3. Treat ceremony scripts as security-critical software. Review them, test them on non-production keys, pin and record versions, and keep logs. A rehearsal, which is how the report describes the April 2025 round trip, must not be able to produce objects that flow into production.

4. Make object identity unambiguous. Near-identical labels are a human-factors risk. PKCS#11 does not require CKA_LABEL or CKA_ID to be unique, and version 3.0 of the standard added CKA_UNIQUE_ID, assigned by the token.1112 Use distinct, structured labels, record object handles and unique identifiers in the ceremony record, and have the CSR step assert that it is using the intended object.

5. Verify migrated keys independently. At a minimum:

  • compare the public key before and after migration, byte for byte;
  • have the destination key sign known test data that is verified with the original public key;
  • run key-quality checks, for example testing an ECDSA public key against small multiples of the base point, a check Chrome says common linters did not include before this incident;7
  • keep HSM-side generation evidence and attestation where the product provides it.

6. Verify what you publish. Validate every CRL signature against the issuing certificate’s public key before publication, automatically, and treat a failure as an incident. A CRL that does not verify is evidence that two different private keys may be in play.

7. Use independent assurance for rare, high-risk operations. A witnessed ceremony shows a script was followed; it does not show the script was right. For a one-off migration, someone who did not write the procedure should review the design, the code and the evidence trail, and verify the outputs.

Reading the standards carefully

The incident report cites “TLS Baseline Requirements §6.2.6” and “CPS §6.2.6”. We read the current Baseline Requirements text. Section 6.2.6 there concerns encrypting a private key for transport when an Issuing CA generates a key on behalf of a Subordinate CA, and section 6.1.1.1 requires a key-generation script, witnessing or recording, and controls without naming a key-strength test.13 So the host-memory conclusion in the report rests on eMudhra’s own CPS, which we have not reviewed. Section 6.2.7 requires CA keys to be protected in a device validated to FIPS 140-2 level 3, FIPS 140-3 level 3 or a Common Criteria profile.13 Community participants have asked whether this is relevant to host-side key handling;10 the report does not address it, and we express no view.

Controls checklist for an HSM migration

Use this before migrating CA or other high-value keys between HSMs.

  • ☐ Migration method approved, with a documented reason if it departs from the vendors’ native mechanisms
  • ☐ Both vendors’ documentation reviewed, and unsupported steps identified
  • ☐ Every script reviewed by a second person and tested on non-production keys
  • ☐ No private key value held outside a validated module, or an approved and assessed exception recorded
  • ☐ Unique labels and identifiers; the ceremony log records object handles and unique IDs
  • ☐ Public keys compared before and after migration
  • ☐ Known-answer sign and verify test on the migrated key
  • ☐ Key-quality checks recorded, including small-scalar checks for elliptic-curve keys
  • ☐ CSR generation asserts the intended object
  • ☐ CRL signatures validated automatically against the issuing certificate before publication
  • ☐ Rollback and continuity plan for issuance, CRL and OCSP
  • ☐ Independent review of design, scripts and evidence before key activation
  • ☐ Inventory of all CA keys and their expiry dates

What this article does not claim

We have not seen eMudhra’s CPS, the ceremony script or its logs. The incident report is eMudhra’s account and the independent review is not finished. We make no allegation about any individual or organisation, and the report itself says the HSMs were not at fault. If the record changes, we will update this article and note the change.

Next step

SafeCipher plans and reviews HSM and PKI key migrations independently of any manufacturer. If you want a second pair of eyes on a migration design, ceremony scripts or the evidence trail, start with a cryptographic audit, or read about our HSM migration services, PKI and HSM migration and HSM architecture services. If you are comparing HSM platforms before a move, our vendor-neutral HSM comparison guide sets out the questions to ask, and our enterprise PKI architecture page covers CA hierarchy design.

References

All sources accessed on 16 November 2026.

  1. eMudhra, “Full Incident Report”, Mozilla Bugzilla 2083449, comment 82, 10 Nov 2026: bugzilla.mozilla.org/show_bug.cgi?id=2083449#c82
  2. Mozilla Bugzilla 2083449, comments 0 and 5 (Mozilla timeline and key analysis), and comment 13 (Sectigo revocation): bugzilla.mozilla.org/show_bug.cgi?id=2083449
  3. Mozilla Security Blog, “Protecting Firefox users following the compromise of eMudhra intermediate certificates”, 20 Oct 2026: blog.mozilla.org/security/2026/10/20/…
  4. “Compromised eMudhra Intermediate Certificates”, Mozilla dev-security-policy thread, 20 Oct 2026: groups.google.com/a/mozilla.org/g/dev-security-policy/c/0R3XLJL6FuE
  5. Utimaco, “Compromised private keys in eMudhra certificates”, 23 Oct 2026: utimaco.com/news/blog-posts/compromised-private-keys-emudhra-certificates
  6. eMudhra, “Statement on Recent Certificate Security Incident” (status page, Update 7, 30 Oct 2026): emudhra.com/en/post/certificate-security-incident
  7. Chrome Root Program, Bugzilla 2083449, comment 50, 27 Oct 2026: bugzilla.mozilla.org/show_bug.cgi?id=2083449#c50
  8. eMudhra, Bugzilla 2083449, comment 11, 19 Oct 2026: bugzilla.mozilla.org/show_bug.cgi?id=2083449#c11
  9. Mozilla Bugzilla 1970728, eMudhra CRL signature issue (2025): bugzilla.mozilla.org/show_bug.cgi?id=1970728
  10. Public questions on the incident, Bugzilla 2083449, comments 85 to 91 (10 Nov 2026): bugzilla.mozilla.org/show_bug.cgi?id=2083449#c85
  11. OASIS, PKCS #11 Cryptographic Token Interface Base Specification v3.0, section 4.4 (CKA_UNIQUE_ID): docs.oasis-open.org/pkcs11/pkcs11-base/v3.0
  12. Securosys, “Identifiers and Labels in PKCS#11” (uniqueness of CKA_LABEL and CKA_ID): docs.securosys.com/pkcs/Concepts/identifiers
  13. CA/Browser Forum, Baseline Requirements for TLS Server Certificates, sections 6.1.1.1, 6.2.6 and 6.2.7 (text read 16 Nov 2026): cabforum.org/working-groups/server/baseline-requirements/requirements

Leave a Reply

Your email address will not be published. Required fields are marked *