Now domain-agnostic — configure any practice with Domain Packs (immigration is the reference pack)

Security & Compliance

Audience: Security reviewers, compliance officers, and engineers configuring access control or encryption.
What you will accomplish: Understand the security model — audit logging, RBAC, privilege gates, encryption — well enough to evaluate, configure, or extend it.
Prerequisites: Read Architecture first; familiarity with AES-256-GCM and OAuth 2.0 helps for the encryption page.
Estimated time: ~5 minutes (this page); ~25 minutes for the full section.

Security is not a feature — it is the foundation. Every design decision in the Case Automation MCP Server is made under the assumption that a missed deadline, a privilege breach, or an incorrect external communication is a malpractice liability.

Where to start

Read in this order for a complete security picture:

  1. Audit & Logging — The tamper-evident hash chain that records every action
  2. RBAC & Privilege — Who can do what, and the non-overridable privilege gate
  3. Encryption & Secrets — How data is protected at rest and in transit

Security scorecard

DimensionScoreGrade
Confidentiality / privilege96A
Auditability96A
Access control93A
Encryption95A
Secret hygiene92A−
Observability security90A−

Overall security posture: 96 / 100

Key security invariants

  1. Every write audited before complete. Hash-chained, append-only, tamper-evident. No action is "done" until the audit record is written.
  2. Default-deny RBAC. The agent service identity has the minimal grant set. Every tool and matter is scoped to roles.
  3. Privilege gate is non-overridable. A Document.privileged=True can never reach an external recipient, regardless of approval tokens.
  4. PII never in logs. A shared scrubber redacts A-numbers, SSNs, passport numbers, DOB, email, and phone across all observability signals.
  5. No credentials in source. All secrets read from the secret store at runtime. OAuth refresh tokens stored encrypted in Postgres.
  6. Idempotency prevents double-sends. Every external action carries an idempotency key. Retries cannot double-send or double-file.

Human-in-the-loop gates

External and irreversible actions require explicit human approval. This is a default-on safety invariant, not a configuration option.

Action tierGate required?DefaultOverrideable?
Read (no side effects)NoN/AN/A
Write (reversible)One-click confirmYesPer-workflow config
External / irreversibleExplicit approval, multi-channelAlways onNo

Gate channels:

  • MCP — The agent client surfaces an approval prompt
  • Web — Sidecar serves an approval page at /approvals/{token}
  • Email — Approval link sent to the required role's inbox

All three channels share the same token pool and enforce single-use atomicity.

Compliance path

The system is designed toward SOC 2 Type II with GDPR-awareness:

  • Data residency configurable per CAM_RESIDENCY_REGION
  • PII scrubber across logs, traces, and metrics
  • Audit log exportable for review
  • RBAC with least-privilege principle
  • Encryption in transit (TLS) and at rest (AES-256-GCM)

HIPAA path is possible if required by the firm's practice area.

Next steps


Last updated: 2026-06-01