
Compliance
SOC 2 Reports: Reading a Vendor's Exceptions and CUECs
SOC 2 reports are conditional: a clean opinion can sit beside a test that failed 25 of 25 samples. How to read scope, exceptions, CUECs and carve-outs.
A vendor sends its SOC 2 report under NDA. Procurement files the PDF, someone ticks "SOC 2: yes" on the vendor record, and nobody reads past the opinion page. That discards most of what the report says.
This is the consumer side; the producer side is in SOC 2 on self-hosted Kubernetes. Here you hold someone else's report and must decide what it lets you rely on.
A SOC 2 Report Is a Conditional Opinion, Not a Badge
The worked example throughout is a real, publicly posted document: the 2025 Type 2 SOC 2 report of ANS, a UK managed-services provider. It is a Type 2 examination performed under AT-C 105 and AT-C 205 covering 1 August 2024 to 31 July 2025. A newer report may have replaced it; it is used because it is public and complete, not to comment on the vendor.
Its opinion is unmodified. The auditor states that, "in all material respects," the controls "operated effectively throughout the period" to provide reasonable assurance that the service commitments were achieved — and then finishes the sentence: "if complementary subservice organization controls and complementary user entity controls assumed in the design of ANS' controls operated effectively throughout that period."
That trailing clause is the whole article. The opinion holds on written conditions: one sits with the vendor's suppliers, the other with you.
Section by Section: Where the Conditions Hide
Section numbering is not standardised; firms order the parts differently, so navigate by the table of contents. In the example report the order is:
- Management's assertion. The vendor's own signed statement, dated 13 October 2025 — about ten weeks after the period ended. Its date is the first data point for the gap you will measure later.
- Independent service auditor's report. The opinion, its basis, and the restricted-use paragraph. Use is restricted to the vendor, current and prospective user entities and business partners, their practitioners, and regulators — readers "who have sufficient knowledge and understanding" of matters such as the nature of the service, the complementary controls, and the criteria.
- The system description. Written by management, opined on by the auditor. Beyond the boundary and components it carries subsections headed "Changes to the System Since the Last Review," "Incidents Since the Last Review" and "Criteria Not Applicable to the System," plus the subservice organisations and the complementary user entity controls. Read all of them; scope narrows quietly there.
- Tests of controls and results. For each criterion: the control as the vendor specifies it, the test the auditor applied, and the result. This is the audited evidence.
- Other information provided by the service organization. Here, management's responses to testing exceptions. The auditor's report says this section is "not a part of the description" and that management's response "has not been subjected to the procedures applied in the examination."
Section 4 is what the auditor found. Section 5 is what the vendor says about it. Only one is assured.
The Opinion Paragraph: Unmodified, Qualified, Adverse, Disclaimer
AT-C 205, as amended by SSAE 21, defines a modified opinion as "a qualified opinion, an adverse opinion, or a disclaimer of opinion." Each has recognisable wording, and the opinion paragraph is the one place worth reading word by word:
- Qualified (AT-C 205.72–.73). Two routes lead here: the auditor has the evidence and finds misstatements "material, but not pervasive," or cannot obtain sufficient appropriate evidence and judges the possible effects of undetected misstatements "could be material but not pervasive." The tells differ: "except for the effects of the matter or matters giving rise to the modification" for a known misstatement, "except for the possible effects of the matter(s)" for a scope limitation. Check whether the basis paragraph touches anything you rely on.
- Adverse (.74–.75). Misstatements are "both material and pervasive." The opinion states that the subject matter "is not in accordance with (or based on) the criteria, in all material respects" — and in SOC 2 the subject matter is the description and the design as well as operating effectiveness.
- Disclaimer (.77–.78). The auditor could not obtain sufficient appropriate evidence and the possible effects "could be both material and pervasive." The report states that the auditor "does not express an opinion on the subject matter." There is no opinion to rely on at all.
- Unmodified. None of the above. Not the same as "no exceptions."
A qualified opinion needs either a known misstatement the auditor judges material but not pervasive, or a scope limitation whose possible effects could be material but not pervasive; material and pervasive moves the report to adverse or disclaimer. Exceptions found in testing that the auditor does not judge material leave the opinion unmodified and show up only in the results. The opinion page cannot tell you which controls failed; Section 4 can.
An adverse opinion or a disclaimer gives you no assurance; the decision is escalation or exit. The comparison with an ISO certificate — which has no opinion paragraph at all — is covered in ISO 27001 vs SOC 2.
Scope Before Results: Categories, Boundaries and the Period
First establish what the opinion is about:
- Which trust services categories? The example report covers the criteria "relevant to Security" and nothing else. It gives no opinion on the Availability or Confidentiality categories, even where a continuity-related common criterion was tested. If your risk sits in uptime, it is the wrong document, not a weak version of the right one.
- Which system? The opinion covers the system the description names. Match the system name and boundary to the product, region and tier you buy.
- Which period? A Type 2 opinion covers a stated window. The period has to overlap the window in which you depended on the service, or it describes someone else's history.
Then read "Criteria Not Applicable to the System" and "Changes to the System Since the Last Review." A criterion declared not applicable is one the auditor did not test. A large change mid-period means part of the period was a different system.
Reading Section 4 and the Exceptions Table
Readers skim Section 4 for "No exceptions noted." Search it instead for the language of a failed test. In the example report one control — personnel acknowledging the employee handbook and code of conduct annually — was tested with three procedures. Two found no exceptions. The third did not:
# EXCERPT — layout paraphrased from the example report's Section 4 table;
# only the Test Results sentence is verbatim. One control, three procedures.
Criteria: CC1.1, CC1.5, CC2.2 (the same three procedures repeat under
each; CC1.0 is a group label in the criteria column, not a criterion)
Control activity: Personnel are required to acknowledge the employee
handbook and code of conduct on an annual basis.
Procedure 1 (inquiry): Inquired of the Head of Compliance -> No exceptions noted.
Procedure 2 (inspection): Inspected the employee handbook -> No exceptions noted.
Procedure 3 (inspection): Inspected signed acknowledgements for a
sample of current employees -> "Testing of the control
activity disclosed that the employee handbook and code of conduct
were not acknowledged for 25 of 25 current employees sampled."
Copy into the register: control, criteria, result, relied_on, conclusion, owner.The auditor's own test found the control not operating for 25 of 25 employees sampled. The opinion is unmodified, "in all material respects." The report does not explain the materiality judgement, and nothing here suggests it was wrong — but it shows why the opinion page alone is not a review.
For each exception, answer three questions and write the answers down:
- Do I rely on this control? A missed code-of-conduct acknowledgement matters less to you than a failed access-revocation test on accounts holding your data. Relevance is your call, not the auditor's.
- Is the failure a sample or a population? "3 of 25" and "25 of 25" are different findings. The second means every sampled item failed.
- What does management say, and is it assured? The vendor's response sits in Section 5 and was not examined. Treat it as a claim to verify next cycle, not as remediation.
Deloitte's 2025 practitioner guide on reviewing third-party SOC reports recommends this: "create a matrix of testing exceptions and document your conclusions." The register below is that matrix in a form you can diff.
CUECs: Obligations the Report Moves Onto You
Complementary user entity controls are controls the vendor's design assumes its customers operate. The opinion is explicitly conditional on them, and the auditor is equally explicit about what was not checked: "Our examination did not include such complementary user entity controls and we have not evaluated the suitability of the design or operating effectiveness of such controls."
So every CUEC is a control in your environment that the vendor's assurance depends on and the service auditor did not test them. The example report's list includes: "User entities are responsible for developing their own disaster recovery and business continuity plans that address the inability to access or utilize ANS services." Your plan for losing the service is your obligation, in writing.
The report also says the list is incomplete: these items "represent only a part of the control considerations that might be pertinent at the user entities' locations," and user entities' auditors "should exercise judgment." Deloitte's guide makes the same point from the reader's side — "There may be other CUECs not suggested by the SOC report" — and recommends a matrix of all CUECs with the rationale for their relevance.
The practical test for each CUEC: name the control you run, the person who owns it and the evidence that it operated. A CUEC with no owner is a gap the vendor's report has documented for you.
Carve-Outs and the Gap Since Period End
The example report carves out its cloud hosting provider, and the auditor says what that means without softening it: "The description does not disclose the actual controls at the subservice organization. Our examination did not include the services provided by the subservice organization, and we have not evaluated the suitability of the design or operating effectiveness of such complementary subservice organization controls." Each carved-out subservice organisation needs its own report in your file, read the same way, or that layer of your dependency has no assurance at all.
The carved-out provider is named in the description: Microsoft Azure. The description then lists the complementary subservice organization controls the design assumes Azure operates, including physical-access controls mapped to CC6.4 and CC7.2. Those are the same kind of condition as a CUEC, pointed at the vendor's supplier instead of at you. This examination did not evaluate them. Physical security of the data centres is part of the Security story the opinion tells, and the evidence for it lives in a different report, issued by a different auditor, on a different schedule.
So the reading task doubles. Obtain the provider's own report, check that its period overlaps the vendor's, confirm that the controls the vendor's description assumes are actually in its scope, and read its exceptions and its own carve-outs the same way. The chain stops only where a report covers a layer inclusively or where you decide the residual risk is acceptable and write that decision down.
Some sections will be quiet, and quiet is still a result. In the example, "Changes to the System Since the Last Review" and "Incidents Since the Last Review" both say nothing significant occurred, and "Criteria Not Applicable to the System" says all common criteria for Security were applicable. Record those answers too: next year's report will say something, and the register is how you notice the difference.
Then measure the gap. The example report's period ended 31 July 2025 and its assertion is dated 13 October 2025, so it was ten weeks old the day it was issued. Deloitte calls the uncovered stretch the "bridge period" and warns that "A long bridge period can lead to increased risks and reduced assurance." The usual remedy is a bridge letter, which the same guide defines as a letter "from the service organization" indicating whether controls have materially changed since the period. Record who wrote it: it is the vendor's statement, not a second opinion from its auditor.
A Vendor-Report Register You Can Diff
These decisions have owners and dates; they belong in version control. One YAML file per vendor, changed by pull request, gives history, review and a diff each cycle. The vendor below is fictional:
# vendor-reports/example-ticketing.yaml — one file per vendor, reviewed in a PR
vendor: Example Ticketing Ltd # fictional
service: Hosted customer-support ticketing
required_categories: [security, availability, confidentiality]
report:
kind: soc2-type2
service_auditor: Example Assurance LLP # fictional
period_start: 2025-04-01
period_end: 2026-03-31
report_date: 2026-06-12
categories_in_scope: [security, availability, confidentiality]
opinion: unmodified # unmodified | qualified | adverse | disclaimer
file: evidence/example-ticketing/soc2-2026.pdf
bridge_letter:
issued_by: service-organization # a management letter, not an opinion
date: 2026-10-06
covers_through: 2026-09-30
material_changes_reported: false
subservice_organizations:
- name: Example Cloud Hosting # fictional
method: carve-out
service: IaaS hosting for the production environment
report_on_file: evidence/example-cloud/soc2-2026.pdf
exceptions:
- control: vendor-ctl-leaver-revocation # illustrative label, not a TSC criterion ID
result: Access not revoked within the stated window for 3 of 25 leavers sampled
relied_on: true # agent accounts depend on vendor-side revocation
management_response_audited: false
conclusion: >-
Compensated: our IdP disables the SSO identity on the leaver's last day,
and the vendor has no local passwords. Re-test in the next report.
owner: identity-platform
- control: vendor-ctl-change-approval # illustrative label, not a TSC criterion ID
result: Change approval not evidenced for 2 of 40 changes sampled
relied_on: false
management_response_audited: false
conclusion: Not relied on; noted for the next cycle.
owner: platform-security
cuecs:
- text: User entities provision and deprovision their own agent accounts.
source: report
our_control: IAM-07 SCIM provisioning from the IdP
owner: identity-platform
evidence: evidence/iam-07/scim-sync-log-2026Q3.json
- text: User entities maintain DR plans covering loss of the service.
source: report
our_control: BCP-04 ticket intake falls back to the self-hosted mail queue
owner: support-engineering
evidence: "" # not yet exercised — the checker will flag this
- text: Export of all tickets is tested before each renewal.
source: ours # a CUEC the report did not list
our_control: EXIT-01 quarterly full export restored into the fallback store
owner: support-engineering
evidence: evidence/exit-01/export-restore-2026-09.log
review:
reviewed_on: 2026-10-20
next_review: 2027-07-01
decision: accept-with-compensating-controls # accept | accept-with-compensating-controls | escalate | exitA register nobody checks decays. The script below runs in CI over the directory and fails the job when an entry needs a human — a modified opinion, a missing category, an uncovered gap, an unowned CUEC. The gap threshold is engineering judgement, exposed as --max-gap-days; the one-year staleness check is hardcoded, so edit it if your appetite differs.
#!/usr/bin/env python3
"""Flag vendor SOC 2 register entries that need a human decision.
Usage: python3 check_vendor_reports.py vendor-reports/ [--max-gap-days 120] [--today 2026-11-16]
Exits 1 if any entry has findings, so it can gate a CI job.
"""
import argparse
import datetime as dt
import pathlib
import sys
import yaml
OPINIONS = {"unmodified", "qualified", "adverse", "disclaimer"}
def as_date(value):
if isinstance(value, dt.date):
return value
return dt.date.fromisoformat(str(value)) if value else None
def check(entry, today, max_gap_days):
findings = []
report = entry.get("report") or {}
opinion = report.get("opinion")
if opinion not in OPINIONS:
findings.append(f"opinion not recorded as one of {sorted(OPINIONS)}")
elif opinion != "unmodified":
findings.append(f"opinion is {opinion}: read the basis paragraph and escalate")
missing = set(entry.get("required_categories", [])) - set(report.get("categories_in_scope", []))
if missing:
findings.append(f"categories you need but the report does not cover: {sorted(missing)}")
period_end = as_date(report.get("period_end"))
if period_end is None:
findings.append("report period_end missing")
else:
bridge = entry.get("bridge_letter") or {}
covered = max(period_end, as_date(bridge.get("covers_through")) or period_end)
if (today - period_end).days > 365:
findings.append(f"report period ended {period_end}: request the current report")
if (today - covered).days > max_gap_days:
findings.append(f"nothing covers {covered} to {today}: request a bridge letter")
if bridge.get("material_changes_reported"):
findings.append("bridge letter reports material changes: review them")
for sub in entry.get("subservice_organizations", []):
if sub.get("method") == "carve-out" and not sub.get("report_on_file"):
findings.append(f"carved-out {sub.get('name')}: no report on file")
for exc in entry.get("exceptions", []):
if exc.get("relied_on") and not exc.get("conclusion"):
findings.append(f"exception on relied-on control {exc.get('control')}: no conclusion")
if not exc.get("owner"):
findings.append(f"exception {exc.get('control')}: no owner")
for cuec in entry.get("cuecs", []):
for field in ("our_control", "owner", "evidence"):
if not cuec.get(field):
findings.append(f"CUEC '{cuec.get('text', '?')[:50]}': missing {field}")
next_review = as_date((entry.get("review") or {}).get("next_review"))
if next_review is None or next_review < today:
findings.append("review overdue or not scheduled")
return findings
def main():
parser = argparse.ArgumentParser()
parser.add_argument("paths", nargs="+", type=pathlib.Path)
parser.add_argument("--max-gap-days", type=int, default=120)
parser.add_argument("--today", type=dt.date.fromisoformat, default=dt.date.today())
args = parser.parse_args()
files = []
for path in args.paths:
files.extend(sorted(path.glob("*.yaml")) if path.is_dir() else [path])
failed = 0
for file in files:
entry = yaml.safe_load(file.read_text())
findings = check(entry, args.today, args.max_gap_days)
status = "FLAG" if findings else "ok"
print(f"{status:4} {entry.get('vendor', file.name)}")
for finding in findings:
print(f" - {finding}")
failed += bool(findings)
return 1 if failed else 0
if __name__ == "__main__":
sys.exit(main())On the example, python3 check_vendor_reports.py vendor-reports/ --today 2026-11-16 prints FLAG Example Ticketing Ltd and the missing DR evidence. Re-run it with --today 2027-02-16 and a second finding appears: nothing covers 2026-09-30 to 2027-02-16, because 139 days exceeds the default 120. The message starts at the bridge letter's own covers-through date, so read it as "no assurance after this day."
Know what the checker does not do, so nobody mistakes a green job for a review. It never opens the PDF: every field it reads was typed in by a person who read the report, and a wrong transcription passes. It does not test that the evidence paths exist or that the files behind them show the control operating; a non-empty string satisfies it. It does not read issued_by, so it will not tell a bridge letter written by management from anything stronger — that distinction stays in the YAML comment and in the reviewer's head.
That division of labour is deliberate. The script enforces completeness: every exception has an owner, every CUEC has a control and evidence, every carve-out has a report, every gap has a date. Judgement — whether an exception matters, whether a compensating control really compensates, whether an unmodified opinion is enough — happens in the pull request, where a second person reads the diff and the conclusion together. Reviewers who want stricter gates can add them; the YAML already carries the fields.
If you produce your own SOC 2, criterion CC9.2 reads: "The entity assesses and manages risks associated with vendors and business partners." Its points of focus, as revised in 2022, say the entity "inventories, tiers, and assesses, on a periodic basis, threats arising from relationships with vendors and business partners (and those entities' vendors and business partners) and the vulnerability of the entity's objectives to those threats," and "implements procedures for terminating vendor and business partner relationships based on predefined considerations," which "may include safe return of data and its removal from the vendor or business partner system." Points of focus are guidance, not mandated procedures, but a reviewed, versioned register is one direct way to evidence both. The broader concentration question is in third-party risk when one vendor runs your platform.
Optionality: When the Report Says Leave
Agree the exit triggers before you need them:
- An adverse opinion or a disclaimer. There is no assurance to rely on.
- Repeat exceptions on a control you rely on. One cycle is a finding; two is how the vendor runs.
- A CUEC you cannot operate. If the design assumes a control you cannot run, the vendor's assurance does not extend to you.
- A carve-out you cannot cover. A subservice organisation with no report you can obtain is a hole of unknown depth.
Exit is only a real option if it was designed in. That is why the example register carries a reader-added CUEC for a tested export: the termination procedure CC9.2's points of focus describe is cheap to write and expensive to discover you cannot execute. For a team that self-hosts most of its stack, the few SaaS dependencies left are exactly where carve-outs, CUECs and bridge periods land — and exactly where an exit ramp matters most. The same reading discipline applies to other third-party evidence, from penetration-test reports to DORA's ICT third-party register.
§FAQ/Common questions
Frequently asked
What should I look for first in a vendor's SOC 2 report?
The opinion type and its conditions. Under AT-C 205 a modified opinion is qualified ('except for'), adverse or a disclaimer; anything else is unmodified. Then read the end of the opinion sentence: where the design relies on subservice organizations and customers, the opinion is conditional on complementary subservice organization controls and complementary user entity controls operating effectively — controls the auditor states it did not evaluate.
Can a SOC 2 report have exceptions and still have a clean opinion?
Yes. Under AT-C 205 a qualified opinion follows a misstatement the auditor judges material but not pervasive, or a scope limitation whose possible effects could be material but not pervasive. Testing exceptions the auditor does not judge material leave the opinion unmodified and appear only in the test results. One publicly posted 2025 Type 2 report carries an unmodified opinion while its own Section 4 test found a control not operating for 25 of 25 employees sampled.
What are complementary user entity controls (CUECs)?
Controls the vendor's design assumes its customers operate, such as managing their own users or maintaining their own disaster recovery plans for loss of the service. The opinion is conditional on them, the auditor does not evaluate them, and the report itself may say its list is only part of the relevant considerations. Map each one to a control you run, an owner and evidence.
Is a bridge letter the same as an updated SOC 2 report?
No. A bridge letter is a letter from the service organization stating whether its controls have materially changed since the report period ended. It is the vendor's own statement, not an auditor's opinion, and covers the bridge period the report does not.
Is management's response to a testing exception audited?
Not necessarily. In the example report, management's responses sit in a separate other-information section that the auditor states is not part of the description and was not subjected to the examination procedures. Treat a response as a claim to verify in the next report.
Further reading
- SOC 2 on Self-Hosted Kubernetes: Own the Evidence
- ISO 27001 vs SOC 2: Certificate or Auditor's Opinion
- SOC 1 vs SOC 2: Which Report Your Customer's Auditor Needs
- Third-Party Risk When One Vendor Runs Your Whole Platform
- Vulnerability Assessment and Penetration Testing: The Report
- Self-Hosted DORA: The Case Against Concentrated Risk
Executive Briefing
Thirty minutes to clarify your infrastructure risk
Walk us through your vendor footprint and regulatory constraints. We will tell you honestly where sovereignty creates leverage — and where it does not. No pitch deck. No obligation.