513

days the archive was reachable from the internet, 4 Feb 2025 to 2 Jul 2026, per DigiCert's timeline

6

days of confirmed data access by an outside account, 27 Jun to 2 Jul 2026

123

days from first seeing the stranger accounts to confirming they had read data, 18 Jun to 19 Oct

0

certificates affected, per DigiCert; the account was read only and never reached the CA systems

The network was the authorization

DigiCert filed its full incident report on Mozilla's public certificate authority tracker on 1 Nov 2026, and the sentence that matters most is one DigiCert wrote about itself: authorization for a business system must not be delegated to its network position. For 513 days an internal archive of customer contact details, and for some customers the passport or driving licence they had sent in to prove who they were, sat where the internet could reach it. A 4 Feb 2025 infrastructure migration had moved the system out of a VPN-restricted network segment into an availability zone that was externally accessible, and nothing in the application knew the difference. Its account sign-up endpoint still answered. Its default role still read the archive. Its MFA still applied only to accounts assigned to an application. Those had been the application's rules all along. The VPN had been hiding the fact that they were the only rules.

That is the finding in one line, and it is the structural weakness we meet most often on external engagements. An internal tool is built for a trusted network: authentication sized for staff who have already passed a VPN gateway, authorization generous because everyone who can reach it is presumed entitled, monitoring tuned for a quiet subnet. Then the network changes underneath it and the tool is reachable from anywhere while still believing it is inside. No vulnerability a scanner would recognise was exploited at DigiCert. A registration form did what registration forms do, for a visitor it was never meant to have.

The second lesson is about the review that said no. DigiCert's own staff saw non-DigiCert accounts on the system on 18 Jun 2026 and concluded that, because those accounts were not assigned to any application, they could not reach data. Nine days later an outside account began reading the archive, and it took 123 days from that first sighting to confirm it. "The incident was initially handled as an account-creation issue rather than a data-access event," the report says. The gap between what a permission model is meant to do and what it does when a stranger holds the account is the gap a tester is paid to find, and it is not found by reading the design.

We can read all of this because DigiCert is in every browser's root store, and Mozilla's Root Store Policy, section 2.4, requires CA operators to disclose incidents publicly and keep answering questions until they are resolved. Most breached organisations never publish a word of this. Read it as a free post-incident review of a failure that is probably in your own estate.

What DigiCert disclosed

We keep to the report's words. DigiCert identified unauthorized access to an internal business system used to archive information provided by customers, including business contact information and, in some cases, additional personal information of customer representatives; the preliminary report of 20 Oct 2026 calls it information customer representatives provide to validate their identity. On 4 Feb 2025 the system was moved from a VPN-restricted network segment to a different availability zone, and the migration inadvertently made it accessible externally.

The system uses an account manager that DigiCert also uses for its customer-facing products, running as separate instances. Staff sign in to it with MFA. The account manager has an optional self-service registration flow for customer sign-up on other implementations. That flow does not use MFA and was, in the report's word, inadvertently provided to the business system. Accounts registered through it were not subject to MFA and were assigned read-only mode by default, and those unassigned accounts had access to archived data.

DigiCert deployed an authorization fix on 2 Jul 2026 that stopped the account creation and denied archive access to unassigned accounts, not knowing data had been read. On 19 Oct 2026 it confirmed that an unassigned account had read a subset of the archive between 27 Jun and 2 Jul 2026 and determined this to be a reportable event. Most of the data was business contact information: names, business email addresses, phone numbers and business addresses. A small minority of customer accounts included a customer-provided document such as a driver's license, passport, utility bill or other document with name and address. The account could not write records, act as a validation approver, or issue certificates.

DigiCert internal archive, Mozilla Bugzilla 2084460 High
  • System //

    an internal business system archiving customer contact details and, for some customers, identity documents; separate from the CA, CertCentral and DigiCert ONE
  • Exposed //

    4 Feb 2025 to 2 Jul 2026, after a migration from a VPN-restricted segment to an externally accessible availability zone
  • Way in //

    a self-registration endpoint inherited from a shared account manager; accounts made through it skipped MFA and read the archive by default
  • Data read //

    names, business email addresses, phone numbers, business addresses; for a small minority, a driver's license, passport, utility bill or similar document
  • Access window //

    27 Jun to 2 Jul 2026, read only
  • Confirmed //

    19 Oct 2026; preliminary report 20 Oct; full report 1 Nov
  • Certificates //

    0 affected; no validation, issuance or revocation activity tied to the account
  • Count //

    not disclosed as of 7 Nov 2026, despite repeated questions in the public thread

The report gives no count of accounts, records or customers, only "a small proportion of archived data". Members of the public thread asked for the numbers on 5 Nov and again on 7 Nov, and asked whether the exposed information was used during certificate subscriber validation. DigiCert's reply on 6 Nov restated the data types, said affected customers had been contacted, and said the incident had been reported to law enforcement. The count question stood open as we published, and BleepingComputer, which reported the notifications reaching customers on 5 Nov, had no answer to it either.

What is not in scope matters as much. DigiCert says its CA services, CertCentral and DigiCert ONE were unaffected, and that a review of CA and validation logs found no validation, issuance or revocation tied to the account. The CA systems run on their own secure network segments. That is also the point: the crown jewels were segmented and watched. The archive of the documents people send to a certificate authority to prove their identity was not treated as a crown jewel, and it sat next door.

How a migration became read access to passports

The chain below is built from the report's timeline and root cause analysis, at the level of what each stage achieved. Nothing in it needs a vulnerability, a credential or any skill beyond noticing that an internal-looking hostname answers from the internet and has a sign-up page.

Figure 01

From a VPN segment to a passport scan

  1. 01

    Move //

    An infrastructure migration carries the archive into an availability zone the internet can reach. No application code changes, so no application review happens.

  2. 02

    Register //

    The account manager's self-service sign-up, meant for other products, now answers to anyone. Accounts made this way skip MFA.

  3. 03

    Default //

    A new account has no application assigned, so it is given read-only mode. Read-only still reads the archive.

  4. 04

    Read //

    Over 6 days, 27 Jun to 2 Jul 2026, the outside account reads a subset of archived records: contact details and, for a minority, identity documents.

Result //

Identity documents read on the 513th day of exposure, confirmed 123 days after the accounts were first noticed.

Every stage used behaviour the system was designed to have. The migration changed who could trigger it.

The limits are real. The account was read only, it could not touch a certificate, and nobody has shown the data being used. But the people in a certificate authority's validation records are the people who approve certificate requests for their organisations, and the documents they sent are the raw material for impersonating them. A phishing message to a named certificate contact, from a sender who knows which CA they use and their phone number, is a very good phishing message. The breach notification itself is now a pretext an attacker can borrow.

Why the first review said no access

DigiCert's root cause analysis lists three contributing factors that together describe how a sound-looking design fails quietly. The first is that the system's network accessibility changed without a security review of the resulting attack surface. The report's explanation is that the migration was treated as an infrastructure change rather than a security change, which would have triggered a more detailed review and sign-off. That distinction is the whole incident. A change board asks whether the service will stay up, not whether its authentication was ever designed to face a stranger.

The second factor is monitoring that was never reassessed after the move. The archive was meant for internal access over the VPN, so its alerting matched: nothing for account creation through the external endpoint or for request volume against the internal endpoints. A SIEM alert on blocked access was in place by 2 Jul, the day the fix shipped. Application-level alerts on file access and request volume arrived on 15 Oct, after DigiCert had begun to suspect the data had been read. The logs that eventually reconstructed the access window were there all along. Nobody was watching them for this, because the system was not supposed to be reachable.

The third factor is the one to sit with. The authorization model did not contemplate self-registration. MFA was only enabled when an account was assigned to an application that required it, and a self-registered account with no application could still read data. When the internal team saw the strange accounts on 18 Jun, it reasoned from the model's intent: no application, so no access. The behaviour was different, and it took an investigation that began on 12 Oct to prove it. We made the same argument last week about a VPN appliance whose MFA was never asked because the login took a route the policy did not cover. Here the route was a sign-up form, and the policy was tied to an application assignment the stranger never had. MFA attached to one path or one role protects that path and that role. It does not protect the system.

Figure 02

What the migration changed, and what it did not

BEFORE 4 FEB 2025

  1. Archive reachable only over the VPN

  2. Self-registration flow present, reachable by nobody outside

  3. Staff accounts, MFA enforced through their application

  4. Unassigned accounts read the archive by default

  5. Alerting sized for an internal tool

AFTER

  1. Archive reachable from the internet

  2. The same self-registration flow, now public

  3. Self-registered accounts, no application, no MFA

  4. Unassigned accounts still read the archive by default

  5. The same alerting

Nothing in the application changed. Only where it stood.

The rules were the same on both sides of the line. The migration changed who they applied to.

How to find an internal app that went public

The question for your own estate is not whether you have a system like this. Almost everyone does: the ticketing tool, the HR archive, the back-office instance of a product you also sell. The question is whether any of them has moved since it was last reviewed, and what it does when a stranger reaches it. DigiCert's own action items, read in reverse, are the checklist, and the What To Do list below follows them.

The detection that would have shortened this incident from months to minutes is the one for stage two of the chain: someone registering an account on an internal hostname from an address that is not yours. On a web server or reverse proxy log that is a POST to a registration path, on a host that should only ever see private addresses, from a public one. The rule below is the simplest version. Tune the hostnames and paths to your own, and expect the first run to be a map of systems you did not know were listening.

Figure 03

Account self-registration on an internal hostname from an external address

Sigma

title: Self-registration on an internal hostname from a public address
status: experimental
description: A POST to an account registration path on a hostname that
  should only be reached over the VPN, from an address outside the
  private ranges. This is the stage of the DigiCert chain where an
  outsider gained an account. Tune hostnames and paths to your estate.
logsource:
  category: webserver
detection:
  selection:
    cs-method: 'POST'
    c-uri|contains:
      - '/register'
      - '/signup'
      - '/sign-up'
      - '/users/new'
      - '/account/create'
    cs-host|endswith:
      - '.internal.example'
      - '.corp.example'
  filter_private:
    c-ip|cidr:
      - '10.0.0.0/8'
      - '172.16.0.0/12'
      - '192.168.0.0/16'
  condition: selection and not filter_private
falsepositives:
  - An internal hostname that is meant to accept public sign-up
  - Staff on a VPN that exits through a public address range
  - A proxy that rewrites the client address; use the forwarded header
level: high
A starting point to tune against your own telemetry, not a production rule. Replace the hostname suffixes and registration paths with the ones your internal applications actually use.

Two companion checks are cheaper than any rule. Resolve each hostname you consider internal from outside your network; a name that answers with a public address has already failed. Then look at what a freshly created account with no role can see on each internal application. DigiCert's fix was to make every endpoint check application assignment and role and return 401 otherwise. Our guides to exposed admin and debug endpoints and broken access control testing cover both from the tester's side.

How we test internal systems from the outside

On an external engagement we do not start from the asset list the client gives us. We start from what the internet can see: DNS, certificate transparency logs, cloud provider address ranges and the hostnames that leak through public pages. Certificate transparency is often the richest source, since every certificate issued for an internal-sounding hostname is public. External recon and subdomain enumeration is the first stage of that work, and our DNS records tool is a quick way to see what a single hostname answers to a public resolver. The hostnames that worry us have admin, archive, internal or ops in the name and answer on a public address.

For each of those we ask three questions. Can a stranger get an account, through a sign-up form, a reset flow, a federation endpoint or an API? What does the system give an account with no role and no application? And is MFA a property of the system, or of one path into it? The third question is where most answers go wrong. We get past MFA on most engagements and we almost never break it; we find the path it was not attached to. Authentication and session testing describes how we work through those paths.

We never read customer data on an engagement. Proof that an unassigned account can list the archive, with record counts and field names and nothing else, is enough to put a finding at the top of a report. What DigiCert needed was that proof in February 2025, from someone on the outside. Its own external penetration test came on 31 Aug 2026, after the fixes, and confirmed the door was shut. A test confirms the state of the door. Only the logs, read with the right question, tell you who came through it before.

What is still coming

Expect the count to stay contested. The root program thread asked twice for the size of the exposed data set and whether it was used in subscriber validation, and DigiCert had not answered as we published. Mozilla's policy lets it demand a plan of action with milestones or additional audits, and the thread is public, so anyone can watch what happens next.

Expect the class to recur at certificate and registration authorities, because an archive of identity documents is a byproduct of their business and is rarely guarded like the signing keys; DigiCert's report links two other CA incidents from October 2026 concerning registration authority records. For every organisation that has sent a passport scan to a CA or a registrar, the question is which vendors keep it, where, and behind what.

Expect the pattern itself to keep appearing in your own estate. Cloud migrations are the most common way an internal system changes its network position, and most migration runbooks have a line for connectivity and none for attack surface. The durable fix is the one DigiCert wrote into policy on 28 Aug 2026: any change to the network accessibility of an internal tool requires a security sign-off. The cheaper fix is to make every internal tool enforce its own authentication and authorization as if the network were already gone. One day it will be.

What to do today

  1. Inventory every application whose protection is the network it sits on. Internal tools, back-office instances of customer products, archives, admin consoles. Record what each does when reached by a stranger, not what the design says.

  2. Resolve your internal hostnames from outside and scan what answers. Anything on a public address or a public load balancer is a finding until proven otherwise. DigiCert's last action item was exactly this audit.

  3. Create a throwaway account on each internal application and see what a role-less account can read. If the answer is anything but a 401, fix the authorization model, not the firewall.

  4. Make MFA a property of the system, not of one application assignment or login path. Self-registration, reset and API flows must all land on the same policy.

  5. Alert on account creation and request volume at the application layer, routed to responders. The Sigma rule above is a starting point; an alert that lives only on the firewall does not survive a network change.

  6. Add a security sign-off to every change that alters where an internal system can be reached from. Migrations, new availability zones, load balancer and WAF changes, with one named owner per system, as DigiCert now requires of itself.