Security Posture

Where CMMD is on security, where it's going, and the honest gaps. Customer-shareable. Encryption, multi-tenancy isolation, auth, audit, vendor list, known gaps.

Where we are on security, where we’re going, and the honest gaps. Customer-shareable.

Field	Value
Audience	Customers, enterprise prospects, internal eng
Last security review	2026-05-16
Owner	@aphaiboon

TL;DR

CMMD treats security as product. We’re a small team, not SOC2 certified yet, but our posture is honest and our roadmap is concrete. Sensitive customer data (workspace content, Brain documents, chats) is encrypted at rest and in transit, scoped to organizations by repository-pattern enforcement, and isolated from any AI training data pipeline.

Where we are today

Area	Status	Notes
Encryption at rest	✅ Active	Neon Postgres provides AES-256 at rest
Encryption in transit	✅ Active	TLS 1.3 on all connections
Multi-tenant isolation	✅ Architecturally enforced	Repository pattern — every query org-scoped (see ADR-0002)
2FA / MFA	✅ Available	TOTP supported; required for admin actions; not yet enforced workspace-wide
SSO (SAML, OIDC)	🟡 Planned	Q3 2026 target for enterprise tier
SOC2 Type 1	🟡 Planned	Audit scheduled Q3 2026
SOC2 Type 2	🟡 Planned	12 months after Type 1
ISO 27001	🟡 Considered, not yet committed	Reassess at enterprise revenue threshold
HIPAA	🚫 Not a target	CMMD is not for PHI. Acceptable Use Policy excludes.
GDPR posture	✅ DPA available on request	EU data residency planned Q4 2026
Pen test	🟡 Self-test only	External pen test Q3 2026
Bug bounty	🟡 Informal	Formal program planned post-SOC2
Vulnerability disclosure policy	✅ Published	See Vulnerability Disclosure Policy

Architecture controls

Multi-tenant isolation

Every database table that holds customer data has organization_id
Repository pattern (CI-enforced — scripts/enforce-repository-pattern.sh) blocks new direct db imports
OrgQueryParams makes organizationId required at compile time
0 cross-tenant data leaks since adoption (see ADR-0002)

Authentication

Email + password with bcrypt hashing
Session cookies (httpOnly, secure, SameSite=Lax)
Session token rotation on privilege change
2FA via TOTP (Google Authenticator / 1Password etc.)
Passwordless SSO planned via OAuth/OIDC (Q3 2026)
Magic-link sign-in available

Authorization

Org-scoped roles: Owner / Admin / Manager / Member
Team-scoped membership for filtering (not isolation)
Feature flag gating for staged rollouts (isEnabled(key, orgId, userId))
Audit log capture for admin actions (planned UI; events captured today)

Secrets management

Secrets in environment variables, never committed to repo
Production secrets in Railway’s encrypted env layer
1Password as shared secret vault for human-readable credentials
scripts/check-forge-security enforces sandbox + iframe + Zod validation for Forge

AI / model boundary

Customer data NEVER used for AI training — hard architectural rule
Apex model serving infrastructure is isolated from training data pipeline
Anthropic proxy (current) does not train on API data (per their published terms)
See Model Card for full safety posture

Operational controls

Logging & monitoring

Structured JSON logs to Railway with required stack: on all errors
PostHog for product analytics (event-only — no raw PII in events)
Alerting on auth failure spikes, 5xx error rates, latency outliers (in progress)

Backups & DR

Neon Postgres point-in-time recovery (30 days)
Daily snapshot exports to S3 (planned)
Disaster recovery runbook in Engineering → Runbooks
RPO target: 1 hour. RTO target: 4 hours.

Incident response

On-call rotation documented in Engineering → On-call
Severity definitions (P0/P1/P2/P3) standardized
Postmortem template for any P0/P1
Customer notification commitment: within 24 hours of confirmed customer-impacting incident

Vendor / sub-processor list

(Full sub-processor list maintained on a public page when GA.)

Vendor	Purpose	Data exposed
Railway	Hosting	Encrypted app traffic, logs
Neon	Postgres database	Customer workspace data (encrypted at rest)
Anthropic (current)	LLM provider for Sidekick	Sidekick chat content (per their API ToS — not used for training)
PostHog	Product analytics	Event metadata only — no PII
Composio	Integration broker	OAuth tokens, integration metadata
GitHub	Code repo	No customer data
Mercury	Banking (internal use)	CMMD’s banking — not customer data

Known gaps (honest list)

These are real and we’re not pretending otherwise:

🟡 No SOC2 attestation yet — audit scheduled Q3 2026
🟡 No external pen test on record — scheduled Q3 2026
🟡 No formal bug bounty program — informal disclosure policy active
🟡 SSO/SAML not yet shipped — Q3 2026 target
🟡 Audit log UI not yet customer-visible — events captured, surface TODO
🟡 Single-region (US-East) — EU data residency planned Q4 2026

If a prospect requires any of these today, we’ll be honest and they should wait for the relevant milestone or use a different tool until then.

What we will NOT do

🚫 Train models on customer workspace data
🚫 Sell customer data to third parties
🚫 Add adtech / tracking pixels in product surfaces
🚫 Hide a breach — customer notification commitment is 24h for confirmed customer-impacting incidents

Reporting a security concern

Concern	Where
Suspected vulnerability	security@cmmd.ai — see Vulnerability Disclosure Policy
Suspected data exposure	security@cmmd.ai — P0 response
Account takeover suspected	support@cmmd.ai + revoke sessions in Settings
Other concerns	Trust & Safety form (planned)