InfraSage Trust Center
Everything a procurement or security team needs to evaluate InfraSage, in one place. If something you need is missing, email [email protected]; we reply within 1 business day.
Deployment posture
InfraSage supports three deployment models. In every one of them your telemetry and RCA outputs stay under your control, and we don't backhaul customer data to a shared analytics warehouse.
| Model | Where InfraSage runs | Where your telemetry lives | Where LLM calls go |
|---|---|---|---|
| BYOC (Bring Your Own Cloud) | Your K8s cluster | Your ClickHouse (never leaves your VPC) | Your AWS Bedrock account (Anthropic Sonnet/Haiku) |
| Self-hosted | Your infra | Your ClickHouse | Your Bedrock / Anthropic / OpenAI / local |
| Managed SaaS (not yet available) | InfraSage-managed AWS | InfraSage-managed ClickHouse (single-tenant DB per customer) | Anthropic direct on your behalf |
There is no cross-tenant data pooling by default. Contributing fault signatures across tenants is opt-in per tenant and governed by a separate DPA amendment.
Encryption
- In transit: TLS 1.3 for all external ingest (OTLP, webhooks, API); mTLS supported for gateway↔engine and engine↔ClickHouse
- At rest: ClickHouse data on EBS
gp3/io2with AWS-managed KMS encryption by default; customer-managed KMS supported - Secrets: helm chart supports
ExternalSecretsoperator pointing at AWS Secrets Manager or HashiCorp Vault; hardcoded credentials are not permitted in production deploys
PII handling
InfraSage runs an active PII-redaction pass on log payloads before they enter the RCA agent's context:
- Redacted patterns: email addresses, credit card numbers, phone numbers, SSNs, JWT tokens, API keys, session cookies,
Authorization:headers - Location:
internal/redaction/pii.go - Never sent to LLM: raw log lines matching redaction patterns are replaced with
[REDACTED:kind]markers before any prompt construction
For domain-specific patterns (customer IDs, order numbers, custom tokens), configure additional regex rules via REDACTION_CUSTOM_PATTERNS.
LLM data-handling
- Model providers: Anthropic Claude Sonnet/Haiku via AWS Bedrock (default), or direct Anthropic/OpenAI/local via
LLM_PROVIDER - Data retention at provider: AWS Bedrock does not train on customer data or retain it past the invocation window (per AWS Bedrock Data Privacy)
- Prompt caching: we use Anthropic's 5-minute ephemeral cache for the stable system-prompt + tool-def prefix. Cache is per-tenant, isolated by Bedrock account credential
- Cost containment: hard per-incident $ budget circuit-breaker (
RCA_PER_INCIDENT_BUDGET_USD, default $2) prevents runaway-loop scenarios
Audit trail
Every live-infrastructure action (kubectl-equivalent reads, cloud-API reads) is audited to an immutable ClickHouse table with a SHA-256 payload hash. Rows are queryable via the infrasage_live_infra_audit table.
Coming Q2: WORM (write-once-read-many) audit-log export to S3 Object Lock or Azure Immutable Blob for SOC-2 CC8.1 attestation.
Compliance
- SOC 2 Type II: In progress (observation window opening Q3 2026, target attestation Q1 2027). Auditor: TBD.
- GDPR: DPA template available on request via [email protected]
- HIPAA: Not currently certified. Reach out if you need a compliance roadmap.
- ISO 27001: Not currently certified.
Sub-processors
The full list of third-party services that process customer data is on the Sub-processors page.
Security disclosures + incident response
- Report a vulnerability: [email protected] or the InfraSage Security Vulnerability Disclosure Policy
- Coordinated disclosure window: 90 days from initial report, negotiable for critical infrastructure severity
- Incident notification SLA: For BYOC customers, security incidents affecting the InfraSage platform are notified within 24 hours of confirmation; for self-hosted deployments, we notify via the release-notes channel
RBAC + authentication
- RBAC: Backend-supported today; admin UI ships Q4 2026 (see Security → RBAC)
- Authentication: SAML 2.0 + OIDC + local email/password + API keys with per-key scoping
- Session management: JWT with configurable TTL; refresh-token rotation supported
Pages in this section:
- Sub-processors: every third-party service that touches customer data
- Deployment modes: BYOC vs self-hosted vs SaaS specifics
- Encryption posture: TLS, at-rest, KMS, secret management
- Incident response: how we respond to security incidents and vulnerability reports