Refolk
PlaybookEngineering and open source

The Code-Derived Subprocessor Register: Repo Scan to a Gated, Reconciled List

Produce a subprocessor register where every vendor has file-level evidence, entity, role, and DPA status, reconciled both ways and guarded by a merge-time gate.

21 min readLast reviewed November 16, 2026Read as Markdown

Key takeaways

  • No single evidence source finds every vendor: in a 200-app comparison, static analysis found 4.85 trackers per app that dynamic analysis missed, and dynamic analysis found 2.26 that static analysis missed.
  • A subprocessor register row needs evidence (file and line, egress host, or config), a resolved legal entity, a role, data classes, region, DPA status, and a pointer to the vendor's own sub-processor list.
  • The merge-time gate's lead time comes from your own outbound DPA, where notice periods of 10, 14, or 30 days are common, so the gate must hold production data flow behind a flag until the window has run, not just flag the PR.
  • Publishing an updated list on your website does not notify customers under GDPR Article 28, so the register must record how and when each customer notice was sent.
  • In Refolk's index there are 132 Privacy Engineers in the United States against 34 GRC Engineers, and 84% of a 25-person Privacy Engineer sample works at Meta, Google, or Apple, so outside big tech the register owner usually holds a GRC or engineering title.
  • Two of 11 vendor sub-processor pages in one check rendered their list client-side, so a monitor that does not execute JavaScript sees an empty list and reports no change.

Security or legal has asked which outside companies the product actually sends customer data to, and the answer has to survive an auditor's sample and a customer questionnaire. This playbook is for the engineering lead who owns that answer. It turns repos, config, traffic, and finance records into a subprocessor register with evidence on every row, reconciles it both ways against the list you publish, and wires a gate into CI so a new vendor cannot reach production data before your customer notice window has run.

What a subprocessor register must hold

A subprocessor register is a table with one row per outside company the product sends data to, where each row carries the evidence that the flow exists, the legal entity you contract with, its processing role, the data classes it receives, its region, agreement status, and a pointer to that vendor's own sub-processor list. Without the evidence column it is a list of claims; without the role column it is a vendor inventory, not a subprocessor list.

The requirements come from two directions, and the register has to satisfy both at once.

  • GDPR Article 28. A summary of EDPB Guidelines 07/2020 says the processor contract should include a list of authorised sub-processors, together with the details of their processing activities, locations, and implemented safeguards.
  • SOC 2 CC9.2. Practitioner guidance lists the fields auditors expect: vendor name and primary service, data the vendor handles, criticality, assigned risk tier, contract status and renewal date, and owner. I did not retrieve the AICPA criteria text itself, so treat this field list as practitioner consensus, not the standard's wording.
  • Completeness testing. The auditor will compare your vendor list against your accounts payable records, SSO integrations, and infrastructure configuration. Auditors then sample 5 to 15 vendors and ask for the assessment file for each.

That last point decides the design. If the auditor tests completeness against AP, SSO, and infrastructure, your register has to have already done the same comparison, with the result written down.

Three roles, one test

Every row gets exactly one of three roles. A subprocessor is a separate company that processes personal data on your behalf while you act as a processor for your customers. An independent controller processes personal data for its own purposes and decides how. No personal data means the vendor receives nothing that identifies a person.

The test for processor status comes from EDPB 07/2020: an entity is a processor where it constitutes a separate entity in relation to the controller and processes personal data on the controller's behalf. The EDPB also says not every service provider that processes personal data in the course of delivering a service is a processor, and that role depends on concrete activities in relation to a specific processing. That is why the role is assigned per row, from evidence, and never inherited from a vendor's category.

Assigning a role to each register row

Vendor acts on your instructionsVendor uses data for its own purposes
No personal data, on your behalf
Record the evidence and recheck when payloads change
Subprocessor
Full row, DPA, customer notice before first use
Out of scope
Keep the row with the reason it is excluded
Independent controller
Counsel confirms and you disclose it separately
No personal data in the flowPersonal data in the flow
Role follows two facts about the specific flow - whether personal data moves and whose purposes the vendor serves - not the vendor's product category.

I did not find a regulator-published worked example for code hosts, domain registrars, or model APIs. Treat any role you assign to those as a counsel decision, record who made it, and revisit it when the data you send changes.

Keep risk tier and subservice-organization status in separate columns. CC9.2 treats risk-tier placement and subservice-organization status as distinct calls, and merging them into one field is a finding waiting to happen.

Which evidence sources find which vendors

No single source finds every vendor: static analysis sees code that may never run, runtime evidence sees only code that did run, and neither sees a vendor finance pays for but engineering never wired in. Nobody has published a source-by-source comparison on a server-side SaaS codebase, so the only measured numbers come from mobile apps, and you should read them as direction, not as rates for your product.

StudyScaleFinding
Tracking survey200 appsStatic-only 4.85 vs dynamic-only 2.26 trackers per app
Longitudinal compliance7,741 APKsAt least 85% show more leaks under static analysis
Razaghpanah, NDSS 201840,553 domains20.0% of domains were third-party

The 200-app study is the useful one. Static methods found about 2.1 times as many trackers the other method missed, but dynamic methods still found 2.26 per app that static analysis did not. The authors are careful: without ground truth it is not possible to conclude that one method is better. Static approaches may fail to identify trackers loaded at run time; dynamic approaches miss trackers when they fail to trigger the relevant interactions. Another study concludes that only the union of leaks from static and dynamic analysis profiles an app's behaviour. In a study of 350 apps, 75.8% of calls (n=87,497) were initiated by third-party services or domains, which is a reminder of how much traffic is not yours.

How third parties narrow inside observed traffic

  1. All observed domains
    40,553

    Mobile traffic, not SaaS backends

  2. Third-party domains
    8,099

    20.0% of all domains

  3. Advertising and tracking services
    2,121

    The subset a known list catches

In the NDSS 2018 mobile dataset, a known-tracker list covered only about a quarter of third-party domains, so a list-driven scanner sees a narrow slice of the outside companies present.

The six sources, in scan order

Manifests, environment variable names, CSP connect-src, infrastructure-as-code, and card statements have not been benchmarked against each other anywhere public. What I can give you is what each one proves and how each one lies.

SourceWhat a hit provesHow it lies
Code data-flow scanPersonal data reaches a third-party callSilent on vendors it does not recognise
Package manifestsA vendor SDK is installedInstalled is not called; HTTP-only vendors absent
Host literals and env var namesCode is configured to reach a hostDead config; hosts built at run time
CSP connect-src and IaCBrowser or infra may reach a hostAllowlists outlive the vendor
Egress and DNS logsTraffic actually leftOnly paths that ran in the log window
AP, card statements, SSOYou pay for or log in to a vendorPaid tool may hold no customer data

The sources disagree on where to start. Auditor guidance starts from AP, SSO, and infrastructure; academic work starts from code plus traffic. I scan code first because it is the only source that yields file-and-line evidence and the data classes involved, then use finance and SSO as the completeness test, mirroring what the auditor will do to you.

Who should own the register

The register needs one named owner who can read code and talk contract terms; at the largest tech companies that is often a privacy engineer, but elsewhere it is more likely a GRC engineer or the engineering lead, because the privacy engineering title is concentrated in a handful of employers.

132
People titled Privacy Engineer in the United States in Refolk's index
Against 34 GRC Engineers in the US, a 3.9:1 ratio, and only 12 Privacy Engineers in Germany.
TitleCountryTotalTop employer in sampleTop-3 share of sample
Privacy EngineerUnited States132Meta (13 of 25)84%
Privacy EngineerGermany12Vinted (4 of 12)75%
GRC EngineerUnited States34Reddit, FanDuel, Jones IT (2 each)24%

Source: Refolk's index. Totals and sample counts come from the index; shares and ratios are derived.

The shape matters more than the totals. In a sample of 25 US Privacy Engineers, Meta accounts for 13, Google 5, and Apple 3. GRC Engineers are spread out: no employer holds more than 2 of 25. If you are not at a company of that size, do not wait for a privacy engineer hire to start this work. Assign it to whoever can run a scanner and read a DPA, pair them with counsel, and name them in the register's owner column.

These counts measure job titles, not who actually maintains a register. A Security Engineer or Staff Engineer may own it under another title. If you do want someone who has built one before, as a hire or a short advisory engagement, Refolk lets you ask for that profile in plain English instead of guessing at title strings.

The procedure: repo scan to reconciled register

The procedure runs eight steps in a fixed order: define roles, scan code, sweep config, union runtime traffic, reconcile finance, resolve each vendor, reconcile against your published list, then install the gate. No source publishes how long each step takes, so every duration below is my estimate, not a measurement.

From repo to a gated register

  1. Scan code
    Data-flow scanner output with file locations
  2. Sweep config
    Hosts, env vars, CSP, IaC matched or added
  3. Union runtime
    Egress and DNS hosts explained
  4. Reconcile finance
    AP, cards, and SSO matched or ruled out
  5. Resolve and reconcile
    Entity, role, DPA, both-way diff
  6. Gate
    Required check plus upstream monitors
Each stage adds the vendors the previous one cannot see, and only the last stage keeps the result true after the audit.

Build the register

  1. Define roles and scope
    Engineering lead and counsel write a three-role taxonomy (subprocessor, independent controller, no personal data) mapped to the EDPB two-part test. Done when it is written and the escalation path for ambiguous cases is agreed.
  2. Run a static scan of the code
    Run Bearer CLI with --report privacy or privado scan on every repo that touches customer data. Done when you hold a JSON list of third parties with file locations, tied to a commit hash.
  3. Sweep config for hosts the scanner missed
    Grep host literals, environment variable names, CSP connect-src, and infrastructure-as-code. Done when every host maps to a step 2 entry or is a new row.
  4. Union with runtime egress and DNS logs
    Pull egress and DNS logs across the full retention window and union them with steps 2 and 3. Done when every unmatched host is explained.
  5. Reconcile against finance and SSO
    Finance and GRC compare AP, card statements, and SSO integrations to the register. Done when every paid vendor is on the register or ruled out with a reason.
  6. Resolve entity, role, region, and agreement
    Find each vendor's sub-processor list via legal page, trust center, DPA annex, then mailing list. Done when every row has entity, role, data classes, region, and DPA status.
  7. Reconcile two ways against the published list
    Diff the register against the list you publish, in both directions. Done when every mismatch is dispositioned.
  8. Install the gate and upstream monitors
    Add a required check that fails a PR adding a vendor without a register row, plus one monitor per subprocessor. Done when a test PR is blocked and every subprocessor has a feed, rendered-page monitor, or email route.

Notes on each step

Step 1. Write the taxonomy before scanning, or engineers will classify as they go and every row will need re-litigating. Estimate: an afternoon with counsel.

Step 2. Bearer CLI detects third-party services through an internal database it calls Recipes and identifies over 120 data types; the free version performs single-file data flow analysis. Privado detects more than 110 personal data elements, maps flows to third parties, and writes results to .privado/privado.json; it is listed in NIST's PrivacyEngCollabSpace. Save the raw JSON with the commit hash so the evidence column can cite it. Estimate: the scan is quick; triage is the slow part.

Step 3. This sweep is not documented as a single step by any tool, so treat it as my practice rather than a published method. Its job is to catch what the scanner's vendor list cannot: raw HTTP calls, browser-side calls, and vendors provisioned in Terraform or similar.

Step 4. Use the longest retention window you have. A vendor that only receives a quarterly batch export will not show up in two weeks of logs.

Step 5. This mirrors the auditor's completeness test, so do it before they do. Every AP line and every SSO app is either on the register or carries a one-line reason it holds no customer data.

Step 6. Published lists name the contracting entity; Sentry's names "Amazon Web Services, Inc." and "OpenAI, L.L.C." with addresses. There is no standard location: /legal/subprocessors, /subprocessors, /legal/sub-processors, /gdpr/subprocessors, and /legal/service-providers are all in live use. Check the legal page, then the trust center, then the DPA annex, then sign up for the vendor's mailing list. How many vendors publish a machine-readable list or feed has not been established publicly, so record per row which method you are using.

Step 7. Reconcile both ways. A vendor in code but not on your published list is an undisclosed subprocessor and needs a notice decision now. A vendor on your list but absent from all six sources is either dead weight to remove or a flow your evidence missed; find out which before deleting it.

Step 8. Covered in its own section below, because the timing is where most gates fail.

Register row schema (CSV header plus one example row)
vendor,primary_service,legal_entity,evidence_type,evidence_ref,role,data_classes,region,safeguards,dpa_status,contract_renewal,risk_tier,subservice_org,criticality,owner,vendor_list_location,monitor_method,upstream_change_treatment,on_published_list,customer_notice_sent,disposition,last_verified
Sentry,Error monitoring,(entity from vendor list),code_scan,api/src/errors.ts:42,subprocessor,"user id, email, IP",(region),(safeguards),signed,(date),(tier),(yes/no),(level),(name),legal page,RSS feed,(per our DPA),yes,(date),keep,(date)

Keep one row per vendor per distinct flow; add columns, but do not merge tier and subservice status.

Setting the gate's lead time from the notice window

The gate's lead time comes from your own outbound DPA, not your vendors' DPAs, and in SaaS DPAs that notice period is often 10, 14, or 30 days before the new vendor touches customer data. Your vendors' terms below show the range you will see in practice and what the same clauses look like when you are the customer.

VendorNotice before processingObjection window
OneSignal30 days30 days
Amplitude30 days30 days
Atlassian30 daysDuring notice period
AnthropicReasonable notice15 days
Mistral AIEmail by subscription10 days

Source: the vendors' DPAs and a public compilation of sub-processor terms.

Three clauses are worth reading closely in your own DPA:

  • Notice before processing. Atlassian gives notice at least thirty days before allowing any new sub-processor to process customer personal data. If your DPA reads the same way, the clock runs before the first byte flows, not before the PR merges.
  • Deemed consent. Amplitude's customers are deemed to consent if they do not timely object. That only works if the notice actually reached them.
  • Exit rights. Google Cloud's DPA lets a customer terminate up to 90 days after the notice. A badly timed vendor addition can carry churn risk well after the notice window closes.

If your DPA counts in business days, convert before setting the flag date: 10 business days is roughly 14 calendar days. Record the notice send date and the earliest permitted data-flow date in the register row, so the flag's removal is a checkable fact.

Upstream changes are a separate question. When one of your subprocessors adds its own sub-processor, whether that triggers your customer notice depends on how your DPA is drafted. Have counsel answer it once, and record the answer in the upstream_change_treatment column so nobody decides it ad hoc under deadline.

How this goes wrong

The register fails in predictable ways, and most of them produce a clean-looking result rather than an error. Each failure below has a check you can run locally.

A clean scanner report means the scanner found nothing it recognises, not that nothing is there.
  1. Recipe blind spot. Bearer only flags vendors in its Recipes database, so a raw fetch() call to an unlisted vendor produces a clean report. That is a false negative that looks like a pass. Check: diff the step 3 host list against the scanner output; every host the scanner did not name needs a row.
  2. Regex source matching. Privado's open-source version relies on regular expressions to find personal data sources from their context, so oddly named fields are missed and generic names over-match. Check: hand-trace a sample of flagged flows, and a sample of unflagged calls to known subprocessors.
  3. Unexercised paths. Egress logs only show code paths that ran. A quarterly batch export may never appear in the window you pulled, and one study describes its own runtime findings as only a lower bound. Check: look across the full log retention window and ask each team about scheduled jobs.
  4. Gate scope gap. dependency-review-action only sees manifest diffs. A new vendor called over HTTP without an SDK passes it, and GitHub Actions themselves are not scanned by it. Check: run a host-literal and CSP diff alongside it, and include workflow files in that diff.
  5. Empty list page. In one check of 11 vendor sub-processor pages that resolved, two returned HTML containing no occurrence of the word at all, because the list renders client-side. A crawler without JavaScript reports an empty list and never sees a change. Check: render with a headless browser, or use the vendor's feed where one exists.
  6. Notice by publication. Updating your own page does not notify anyone. Per the EDPB summary, the mere publication of modifications on the processor's website does not comply with Article 28. Check: confirm customers receive an email or a subscription notice, and log the send date.
  7. Tier and subservice conflation. CC9.2 treats risk-tier placement and subservice-organization status as distinct calls. A single "risk" column hides one of them. Check: keep two columns and fill both on every row.
  8. Monitors that miss emails. List monitors watch the published list, not the notice emails vendors send. Some vendors announce by email first. Check: route vendor notice emails to a shared inbox that the register owner triages.
  9. Model API defaulted to no personal data. Classifying by vendor category instead of by flow puts the riskiest new vendors in the out-of-scope bucket. Check: for every AI/ML row, list the data classes in the prompt payload before assigning a role.

When a vendor's page renders empty, has no feed, and the DPA annex is stale, the fastest fix is often a named contact at the vendor who can send the current list. Refolk can find that vendor's privacy or compliance lead from a plain-English request, which beats waiting on a generic support queue.

Wiring the merge-time gate

The merge-time gate is a required CI check that fails any pull request introducing an outbound host or vendor SDK without a matching register row. It works only if it watches more than package manifests, because the most common new-vendor path is a plain HTTP call.

Run two checks together:

  • Manifest check. dependency-review-action's deny-packages option takes package URLs (PURLs); if no version is given, it treats the package as a wildcard. Use it to block SDKs for vendors you have rejected or not yet approved.
  • Host diff check. A script that extracts hosts added in the diff, including workflow files, and compares them to an allowlist exported from the register.
Host-literal gate for CI (bash)
#!/usr/bin/env bash
# Fails the PR if it adds an outbound host that has no register row.
set -euo pipefail
BASE="${BASE_REF:-origin/main}"
ALLOW="compliance/register-hosts.txt"   # one host per line, exported from the register

git diff "$BASE"...HEAD -U0 -- . ':!compliance/' \
  | grep -E '^\+' \
  | grep -Eo '(https?|wss?)://[A-Za-z0-9.-]+' \
  | sed -E 's#^[a-z]+://##' \
  | sort -u > /tmp/new-hosts.txt || true

NEW=$(comm -23 /tmp/new-hosts.txt <(sort -u "$ALLOW"))

if [ -n "$NEW" ]; then
  echo "New outbound hosts without a register row:"
  echo "$NEW"
  echo "Add a register row (entity, role, data classes, DPA status, notice date) before merging."
  exit 1
fi
echo "No unregistered outbound hosts."

Export register-hosts.txt from the register on every change; extend the grep if your CSP lists bare hosts without a scheme.

The gate catches the PR. The feature flag from the previous section catches the deploy. You need both: the check makes sure a row exists, and the flag makes sure customer data does not flow until the notice period has run.

Keeping the register true between audits

Keep the register current with three loops: the CI gate for new vendors, one upstream monitor per subprocessor, and a quarterly review with vendor owners. The only published cadence guidance is advice to revisit the vendor inventory quarterly to catch drift; how fast registers actually go stale has not been measured publicly. The directional evidence points one way: longitudinal research found apps leak personal data to more third-party domains over time.

For upstream monitoring, pick a method per subprocessor and record it in monitor_method:

  • Feed. Some vendors publish one; Sentry offers an RSS feed for changes to its subprocessor list. Prefer this where it exists.
  • Rendered-page monitor. For pages that render client-side, monitor with a headless browser and diff the rendered list.
  • Email route. Subscribe a shared inbox to every vendor's notice list, since page monitors miss emailed notices.

Each quarter, rerun steps 2 through 5 against the current code and logs, diff the result against the register, and treat anything new as a failure of the gate worth explaining.

Before you call the register done

  • Every row cites evidence: a file and line, an egress host, or a config location.
  • Every host in the config sweep and egress logs maps to a row or a written exclusion.
  • Every AP vendor, card charge, and SSO app is on the register or ruled out with a reason.
  • Every row has a resolved legal entity, a role, data classes, a region, and a DPA status.
  • Risk tier and subservice-organization status sit in separate columns on every row.
  • Every AI/ML vendor's role was assigned from the data classes actually sent, and counsel signed off.
  • The two-way diff against the published list has zero undispositioned mismatches.
  • Customer notices were sent by email or subscription, with send dates recorded per change.
  • A required CI check blocked a test PR that added an unregistered host.
  • Every subprocessor has a feed, rendered-page monitor, or routed notice email recorded.
  • The upstream-change column is filled from counsel's reading of your DPA.
  • A quarterly review is on the calendar with a named owner.

Next, pull 5 to 15 rows at random and assemble the file an auditor would ask for on each: the evidence, the DPA, the vendor's list, and the notice record. If any row cannot produce that file in a few minutes, fix that row before the auditor picks it.

Questions practitioners ask

Is a model API like OpenAI or Anthropic a subprocessor?

Usually treat it as one until counsel says otherwise, because prompts tend to carry customer data. Sentry's published list, for example, names Anthropic and OpenAI under AI/ML services. Role depends on what the provider actually does with the data in your specific processing, so record the data classes you send and let counsel apply the EDPB two-part test. Defaulting a model API to no personal data is the most common misclassification I see.

Can I build a subprocessor list by grepping for API keys?

No. Grepping for keys finds vendors you authenticate to, but misses unauthenticated HTTP calls, browser-side calls allowed by CSP connect-src, vendors provisioned in infrastructure-as-code, and tools finance pays for that engineering never wired in. Use six sources in order: code scan, manifests, host literals and env vars, CSP and IaC, egress logs, and finance plus SSO. Then resolve each hit to a legal entity and role.

Does updating our subprocessor page count as notifying customers?

No. A summary of EDPB Guidelines 07/2020 states that the mere publication of modifications on the processor's website does not comply with Article 28. Customers need an actual notice, typically email or a subscription they signed up for, sent early enough to cover the notice period in your outbound DPA. Record the send date per change in the register so you can prove it.

How often should the subprocessor register be reviewed?

Review it quarterly with vendor owners as a floor, and let a CI gate catch new vendors between reviews. No published source measures how fast these lists drift, though longitudinal mobile research found apps leak personal data to more third-party domains over time. The quarterly pass catches what the gate cannot see: contract changes, upstream sub-processor changes, and paid tools with no code footprint.

Do my vendors' new sub-processors require me to notify my customers?

It depends on how your own DPA is drafted, so there is no universal answer. Some DPAs treat a vendor's upstream sub-processor as part of your chain and some do not. Add a register column recording how your DPA treats upstream changes, have counsel fill it once, and route each vendor's change notices into the register so the decision is made deliberately rather than by default.

Should I use Bearer CLI or Privado for the static scan?

Either works as the first pass, and neither is enough alone. Bearer flags only vendors in its Recipes database and the free version does single-file data flow analysis. Privado's open-source version matches personal data sources by regular expression, so oddly named fields slip through. Whichever you run, diff its output against a host-literal and config sweep to catch the vendors it cannot recognise.

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next