Vol VI Book 04 — Federal Compliance — Response Automation: The SOAR Playbook Library
Automated containment and enrichment playbooks that convert a detection into a bounded response
0.1 Draft Proposal — Subject to Review
Developed by the CSI Team (Cloud Services Infrastructure); not yet reviewed or approved by the Agency CIO Office, OIS, or leadership. Date Code: 2026-07-09 15:13 ET
Series Context — Volume VI Book 04. The response half of the detection engine: the playbooks that Book VI-03’s rules trigger. It rebuilds the containment speed that a commercial cloud’s real-time revocation provides, approximated in-boundary to minutes.
1 What This Document Addresses
Detection without response only shortens the investigation, not the exposure. The Volume III SIEM/XDR architecture ends at the alert; this book carries the automation that acts on it. It is the in-boundary substitute for the sub-second Continuous Access Evaluation revocation the boundary removes — and it is honest that the substitute lands at minutes, not sub-second, because a playbook triggered by a scheduled analytic cannot beat the analytic’s cadence.
The design rule for every playbook here is confidence-gated action: high-confidence detections auto-contain; medium-confidence detections enrich and ticket for a human; nothing low-confidence auto-acts. Auto-containment on a noisy rule is its own denial-of-service.
2 The Response Ladder
| Confidence | Action | Human in loop? | Substitutes for |
|---|---|---|---|
| High (e.g. rule 04 session-anomaly with corroboration) | Revoke sessions, confirm-compromised, force re-auth | No — auto | CAE sub-second revocation (→ minutes) |
| Medium | Enrich (identity/device/IP context), open ticket, raise risk | Yes | Vendor auto-remediation |
| Standing control | Bound token lifetime to 60–90 min | n/a — config | CAE exposure ceiling |
3 Deployable Artifacts
| Playbook | Trigger (Book VI-03) | Action | Substitutes for |
|---|---|---|---|
revoke-session |
Token-theft / session-anomaly (rule 04) | Graph revokeSignInSessions + confirm-compromised |
CAE sub-second revocation (→ minutes) |
disable-principal |
High-confidence account compromise | Disable account, block sign-in | Automated account containment |
short-token-lifetime (config) |
Standing control | Bound token lifetime to 60–90 min | CAE exposure ceiling |
enrich-and-ticket |
Medium-confidence detections | Enrich with identity/device context, open ticket (human-in-loop) | Vendor auto-remediation |
3.1 Illustrative playbook — session revocation on a high-confidence session anomaly
The revoke-session playbook is the containment path for Book VI-03 rule 04. It runs entirely in-boundary against the management plane; the containment action is a Graph call, and the whole run is itself logged as an incident action into the evidence contract.
// Logic App / Automation runbook (deployed example) — pseudocode of the revoke-session flow
{
"trigger": "Sentinel incident: rule 'token-theft-session-anomaly'",
"steps": [
{ "guard": "incident.severity == 'High' && entities.Account != null" }, // confidence gate
{ "graph": "POST /users/{account}/revokeSignInSessions" }, // AC-12 containment
{ "graph": "PATCH /identityProtection/riskyUsers/confirmCompromised",
"body": { "userIds": ["{account}"] } }, // raise risk state
{ "ticket": "create IR record, link Sentinel incident, attach entity context" },// IR-4 record
{ "evidence": "emit response-action to the evidence contract (Book VI-07)" } // CA-7 evidence
],
"notes": "Containment latency = detection cadence (minutes), not CAE sub-second. TUNE the confidence gate."
}
The short-token-lifetime artifact is not a playbook but a standing configuration: it bounds the maximum exposure window so that even a missed detection cannot leave a stolen token valid indefinitely. It is the exposure ceiling under which the playbooks operate.
4 Deployment
- Deploy playbooks through the pipeline (Book VI-00); grant the automation identity only the Graph scopes each action needs (least privilege — the revocation identity should not also be able to create accounts).
- Wire each Book VI-03 rule to its playbook via the SIEM automation rules; start every auto-containment playbook in alert-only mode and promote to auto-action only after the triggering rule’s false-positive rate is measured (Vol III Book 06).
- Set the standing token-lifetime bound in the certificate & token identity plane (Vol I Book 03).
5 Closure Necessity
5.1 Closure Necessity — a detection with no response is exposure, not control
IR-4 requires incident handling, not just detection. These playbooks are what convert a Book VI-03 firing into a bounded containment action and a documented incident record — the artifact an assessor asks for. Without them, the detection layer shortens time-to-know but not time-to-contain, and the token-theft window stays open for the full token lifetime.
6 Controls Operationalized
| Control | Title | How this book deploys it |
|---|---|---|
| IR-4/6 | Incident Handling / Reporting | Automated containment + enriched incident records |
| AC-12 | Session Termination | Session revocation + short token lifetime |
| AC-2(13) | Disable Accounts (high-risk) | disable-principal on high-confidence compromise |
7 Honest Limits
- Containment is minutes, not sub-second — bounded by the detection cadence. The
short-token-lifetimeconfig is the standing ceiling that makes that latency tolerable. - Auto-containment amplifies a noisy rule into an availability incident; the confidence gate and the alert-only ramp (Vol III Book 06) are mandatory, not optional.
- The playbooks act only on the agency’s own tenant; a compromised identity federated elsewhere is outside their reach.
8 Cross-References
- Book VI-03 — the detections these playbooks respond to.
- Vol I Book 03 — the token-lifetime standing control (short-lived tokens, refresh-token policy gate).
- Volume III — SIEM/XDR — the incident-workflow architecture.
- Book VI-07 — response actions emit to the evidence contract.
- Vol III Book 06 — false-positive measurement that gates auto-action.
9 Provenance
Draft. Containment latency is honestly minutes, not sub-second. The playbook pseudocode is an illustrative flow to adapt, not a deployable runbook. Products are deployed examples of the SOAR-containment mechanism, which is portable across automation engines.