The Technical Control Set
Lesson 4: The Technical Control Set
This is the lesson engineers actually need, because most SOC 2 pain is technical: not “what does the criteria say” but “what does good look like in my stack, and what will the auditor ask me to prove”. There is no universal control list — a control set is something you design — but the families below are the ones nearly every unqualified report contains. For each, the useful question is not “do we have the tool” but “did the control run every single time, all period, and can I show it”.
1. Identity and access (CC6) — the series that sinks most audits
Access control is the largest Common Criteria series and the most frequently excepted. The bar:
- Central identity provider — SSO for every in-scope system. Local accounts are exceptions you have to justify and track.
- MFA everywhere, with phishing-resistant methods for administrators. Every production console, cloud account, code host, and identity provider must be enrolled; a single unenrolled break-glass admin is a finding waiting to happen.
- Unique identities, no shared accounts. Service accounts are inventoried, own a credential store, and are excluded from interactive login.
- Least privilege and role-based access — documented roles, scoped IAM policies, just-in-time elevation for production rather than standing admin.
- Quarterly user access reviews for in-scope systems, with the reviewer confirming or revoking each entry, and the export retained. This is the classic missed control: three quarters performed, one skipped, one exception.
- Joiner / mover / leaver process with an SLA — access granted only on approved request, changed on role change, and revoked within a defined window (same-day or 24–48 hours) on termination. Auditors sample terminations and compare dates; a two-week lag on one departed engineer can qualify the opinion even if policy and tooling are sound.
- Privileged access management — break-glass accounts documented, credentials rotated after use, session logging on production.
2. Encryption and key management
TLS 1.2+ in transit (1.3 preferred) with no plaintext fallbacks; AES-256 at rest for databases, object storage, and backups; managed KMS with defined rotation and separation of duties so no single engineer can both read ciphertext and extract keys; secrets in a vault, never in git — and secret scanning in CI to prove it. Certificate inventory and expiry monitoring are increasingly tested because outages caused by forgotten certs look exactly like control failures.
3. Network and boundary protection
Segmentation of production from corporate and from development; security groups default-deny; no publicly reachable databases or buckets; WAF and rate limiting on public endpoints; VPN or zero-trust access for administrative paths; documented and reviewed firewall rules. Where you run in a cloud, you inherit physical and environmental controls from the provider — but only if you can show you reviewed their SOC 2 report, which means the review is a control with evidence.
4. Logging, monitoring, and alerting (CC7)
- Centralised, access-controlled logging for authentication, authorization changes, admin actions, and data access — retained per policy (commonly 90 days hot and a year archived), tamper-evident so an engineer cannot delete their own trail.
- Alerting that actually routes to a human, with on-call rotation and documented escalation. Auditors increasingly ask what happened when an alert fired: the answer must be a ticket or incident, not silence.
- Anomaly and infrastructure monitoring tied to the availability criteria if uptime is in scope.
- Incident evaluation: how you decide whether an event is a security incident, who declares it, and where it is recorded.
5. Vulnerability and patch management
Scanning cadence on infrastructure and containers, software composition analysis on dependencies, SAST in CI, and cloud/configuration posture checks. The control the auditor tests is not the scan — it is the remediation SLA: define severity tiers (for example critical in 14–30 days, high in 30–60, medium in 90) and show tickets closing within them. Exceptions must be documented risk acceptances with an expiry date. Add an annual third-party penetration test with findings tracked to closure, and — if you have one — a vulnerability disclosure programme or bug bounty.
6. Change management and secure SDLC (CC8)
Nearly every engineering-led company is strong here and still fails on evidence. The expectations:
- Every production change is traceable to a ticket or pull request describing what and why, and reviewed by someone other than the author.
- CI gates: required reviews, passing tests, SAST/dependency checks, and a recorded approval before deploy. Infrastructure changes go through the same pipeline — console hotfixes are an exception path, not a habit.
- Segregation of duties: the person who writes the code is not the only person who can deploy it unreviewed.
- A documented emergency change process — because you will need one at 2am, and a documented process with after-the-fact review is a control while an undocumented one is a deviation.
- Rollback plan and environment separation (dev/staging/prod) so testing happens before production.
- Threat modelling at design time for significant changes, and an AI-specific review step if you ship models or agents.
# The shape of evidence an auditor will sample for CC8
PR #4821 "add retry to billing webhook" → 2 approvals, tests passed
CI run 9931 → SAST clean, deps scanned
Deploy 2026-09-14T16:03Z by alice → approved by bob
Ticket OPS-118 → linked, closed
Rollback: feature flag billing_retry_v2 → documented
7. Backup, disaster recovery, and continuity (A1, CC9)
Backups configured, encrypted, retained per policy, and stored in a separate account or region with restricted deletion. Then the part everyone skips: restore testing. A backup that has never been restored is an assumption. Define RTO and RPO per service tier, run at least an annual recovery test (quarterly for critical systems), and keep the results — dates, scope, outcome, follow-up actions. The business impact analysis and business continuity plan live here too, and yes, the auditor will ask who signed them and when they were last reviewed.
8. Asset and configuration management
A maintained inventory of hardware, software, and data stores in scope. You cannot scope, monitor, or secure what you haven't listed — and an incomplete inventory is the root cause of “how did that server get into production” findings. Pair it with hardened baselines (CIS or provider defaults) and drift detection. For endpoints: MDM enrolment, disk encryption, screen lock, EDR, automatic patching, and a documented BYOD position.
9. Data handling
Classification scheme that people actually use; retention and disposal rules matching contractual commitments; secure deletion on customer request with a record; production data excluded or masked in non-production environments; documented data residency. If Confidentiality is in scope, the disposal control (C1.2) is tested directly — customer deletion requests, met and logged.
10. Physical and environmental
If you are cloud-native this is largely inherited, and your control becomes the review of the provider's report. If you have an office: badge access with a current access list, visitor logging, server-room or network-closet restrictions, and clean-desk practice. Auditors do observe facilities on request, and remote-first companies should document that no in-scope production infrastructure is on-premise.
| Family | Where it usually breaks | What proves it |
|---|---|---|
| Access | Offboarding SLA misses; one skipped quarterly review; standing admin | Review exports with sign-off, termination vs disable timestamps, role matrix |
| Encryption | Unencrypted legacy bucket; secrets in a repo; no key rotation | Config exports, KMS policy, secret-scan reports |
| Logging | Logs exist but nobody watches; 7-day retention vs 1-year policy | Retention settings, alert configs, linked incident tickets |
| Vulnerability | Scans run, SLAs missed, no risk acceptances | Scan history, remediation tickets with close dates |
| Change | Hotfixes to prod outside the pipeline; self-approved PRs | PR + CI + deploy records for the sampled population |
| Backup/DR | Never restored; DR test “planned” for three years | Restore-test reports with date, scope, outcome |
| Inventory | Shadow IT; unlisted data warehouse copies | Inventory with last-verified date vs cloud asset export |
🧠 Knowledge Check
1. An auditor samples ten terminations from the review period. Nine accounts were disabled within 24 hours; one took 15 days. What is the most likely consequence?
2. Which of these is the actual control under CC8, as opposed to an input to it?
3. Your team runs nightly backups but has never tested a restore. How should this be characterised?