Security

How InventDB protects your data — and what we don't yet have.

A short, honest page. Customer-managed encryption keys, audit log on every action, multi-AZ AWS hosting, no model training on your data, full export and deletion on demand. We also tell you what we don't have today (SOC 2 in progress, no HIPAA BAA, no on-prem) so you can decide before you sign.

What every InventDB instance gets.

01 · Encryption

Per-instance master key, audit-logged and rotatable

Every instance has its own AES-256-GCM master encryption key. The key is generated at provisioning, optionally passphrase-derived via Argon2id, and you can rotate it from the admin console. Revoking the key locks the data. We don't claim "we can't read it" — the running InventDB process must use the key to serve queries — we claim that every key access is audit-logged and that you control rotation. On the roadmap for Business plans: AWS KMS-backed customer-managed keys (CMEK) for centralised rotation, revocation, and key-access audit on AWS-native tooling. Today the key lives in your instance's config; KMS will move it to your AWS account.

02 · Audit log

Every action, every surface, one log

Every change — from the AI Canvas, from MCP, from the REST API, from the SDKs — is recorded with the actor, timestamp, the prompt or payload that triggered it, and the records changed. Pull the log from inside the Canvas, or via the audit_log MCP tool for your SIEM.

03 · AI scope

Role-bounded, no training

The AI assistant inherits the logged-in user's exact permissions and cannot exceed them. Inference is Anthropic Claude Opus 4.7 with the no-training opt-out applied at the API level — your data never enters any model-training pipeline. For the production launch, inference will route through the InventDB AI Gateway to Anthropic Claude on AWS Bedrock; pre-launch it routes through the Gateway to the Anthropic API directly. Either path preserves the same no-training guarantee.

04 · Backup & deletion

Daily backups, 30-day deletion

Daily automated snapshots, multi-AZ replication, 30-day retention. On cancellation: full export of records + attachments + saved templates within 7 days, complete deletion of instance and key 30 days after cancellation, with written confirmation.

Where your data lives.

Each Business Subscriber gets a dedicated InventDB Instance — a dedicated InventDB process running as an ECS task on managed AWS infrastructure, in your chosen region (US-East by default; EU-West and AP on request). Personal subscriptions share the underlying compute pool with strong per-user isolation. In both cases, the underlying EC2 hosts and storage volumes may be shared with other Subscribers; isolation is enforced at the application, encryption, and identity layers. Same model as AWS RDS, Snowflake, or Atlassian Cloud — logical isolation, shared physical hardware, cost-efficient buying power passed to you.

Per-Instance components (Business)

  • Compute: Dedicated ECS task per Subscriber on a shared EC2 fleet (m6i family with Reserved-Instance pricing). Sized per tier, auto-restart on failure. One Subscriber's task cannot read another's process memory or filesystem.
  • Records storage: EBS gp3 with a per-Subscriber folder namespace, encrypted at rest with your per-instance master key (KMS-backed CMEK on the roadmap). Cross-Subscriber filesystem access is impossible at the OS layer (separate ECS task, separate filesystem mount path).
  • Attached files: S3 with a per-Subscriber prefix, encrypted at rest with your per-instance master key. Bucket policy + IAM role bound to your task; no other Subscriber's task can resolve your prefix.
  • Backups: Daily encrypted snapshot, 30-day retention. Backups encrypted with your master key — recoverable only via your active subscription and your key. Cross-AZ replication on the roadmap alongside KMS integration.
  • Network: All traffic over TLS 1.3. Public endpoint via shared ALB with host-based routing to your task; optional IP allow-list (in beta).
  • Authentication: OAuth 2.1 + PKCE. SAML / OIDC SSO on the roadmap for Business plans.
  • AI: Anthropic Claude Opus 4.7 via the InventDB AI Gateway with the no-training opt-out applied. Pre-production-launch: Anthropic API direct. At production launch: AWS Bedrock through the Gateway.
  • Audit: Append-only audit log inside your Instance, exportable via API.

Per-user components (Personal)

Personal subscriptions share an ECS task with other Personal users for cost efficiency. Each user's records and files are encrypted with their own key and stored under their own folder/prefix — cross-user access is impossible regardless of co-residence. Revoking your key locks your data even from us.

  • Compute: Shared ECS task with per-user request authentication and per-user encryption-key context.
  • Records & files: Per-user encrypted folders/prefixes on shared EBS / S3, decryptable only with your key.
  • Auth, audit, AI scope: Same as Business — per-user OAuth identity, per-user audit log, per-user AI usage allocation.

How we isolate tenants on shared infrastructure.

A short, technical answer for the IT lead, the auditor, and the procurement reviewer who are about to ask. We run shared physical infrastructure (the same way AWS RDS, Snowflake, Atlassian Cloud, and most modern SaaS does); isolation is enforced at four layers above it.

Four layers of isolation

  • 1. Process isolation. Each Business Subscriber's InventDB server runs in its own dedicated ECS task. No other Subscriber's task can read process memory, ports, or environment variables. Personal users share a task; isolation moves to layers 2–4.
  • 2. Filesystem isolation. Each Subscriber gets a dedicated folder on the EBS volume mounted to their task; the OS mount path is bound to that folder only. For Personal, each user gets their own folder; the running process enforces per-request folder routing under the user's authenticated identity.
  • 3. Cryptographic isolation. Each Subscriber's records and files are encrypted at rest with their own per-instance master key (Business: dedicated key per instance; Personal: per-user key). Even if a co-resident process could see the bytes, it could not decrypt them — the key isn't accessible to it. KMS-backed CMEK for Business plans is on the roadmap; today the key lives in the instance config and the same isolation property holds.
  • 4. Identity isolation. Every read or write is gated by an OAuth 2.1 token bound to the Subscriber's identity. Tasks run with IAM roles scoped to that Subscriber's S3 prefix only. Cross-Subscriber API calls are rejected at the application layer regardless of network reachability.
Same isolation model as AWS RDS, Snowflake, Atlassian Cloud. If those services are acceptable for your data sensitivity profile, InventDB is acceptable. If you require single-tenant physical hardware (FedRAMP High, certain HIPAA / financial-services regulators, on-prem-only mandates), our current managed-only deployment is not a fit — we'll be straight about that on the sales call.

Who else touches your data.

A short, complete list. We use AWS for hosting and Anthropic for AI inference; nothing else processes your data. We update this list when it changes — existing customers are notified by email at least 30 days before any new subprocessor is added.

SubprocessorPurposeRegionData accessed
Amazon Web Services Compute, storage, networking, monitoring (KMS planned) Hosting your InventDB instance US-East-1 (default); other regions on request Encrypted records, encrypted files, encrypted backups
Anthropic Claude Opus 4.7 inference via the InventDB AI Gateway AI Canvas + agent inference Anthropic API today; AWS Bedrock (same region as your instance) at production launch Your prompt + the rows the AI was asked to operate on (no training, no retention beyond the request)
Razorpay Payment processing · PCI-DSS Level 1, RBI-licensed Subscription billing — cards from US, Europe / UK, India, and worldwide India (with global card acquiring) Billing email, card token, plan and invoice data — never any of your record content
Postmark / SES Transactional email Sign-in links, invoices, scheduled report delivery US Email address, the report HTML you scheduled

Honest about today's gaps.

We're early. Here's what InventDB has now, what's in flight, and what's not on the roadmap. If a missing item is a blocker for your use case, tell us — we'll be straight about whether it's coming.

What we have today

  • AES-256-GCM encryption at rest with a per-instance master key (rotatable from the admin console)
  • TLS 1.3 in transit on every surface
  • OAuth 2.1 + PKCE on every surface (Canvas, MCP, REST, SDKs)
  • Audit log on every action across every surface (queryable via the audit_log MCP tool)
  • Daily encrypted snapshots with 30-day retention
  • No-training opt-out applied at the Anthropic API level (no prompt content retained beyond the request)
  • Role-bounded AI — assistant cannot exceed user's permissions
  • Full data export on demand and on cancellation (CSV + JSON + original files)
  • 30-day instance + key deletion on cancellation, with written confirmation
  • 44 MCP tools (records, attachments, search, audit, bulk-insert, report templates, schema introspection, more)

Roadmap — planned, not promised yet

  • AWS KMS-backed customer-managed keys (CMEK) for Business plans — key in your AWS account, every key access audited via CloudTrail
  • AWS Bedrock routing for AI inference at production launch (today the Gateway calls Anthropic API direct — same no-training guarantee, different surface)
  • Multi-AZ replication + cross-region snapshot
  • SOC 2 Type II — planned; no audit firm engaged yet, no committed date
  • SAML / OIDC SSO for Business plans
  • Customer-supplied IP allow-list (in beta — ask for early access)
  • Tier-segmented support SLAs (1-hour sev-1 acknowledgement on Business) once the team has 24×7 coverage
  • Public status page at status.inventdb.com
  • Webhooks for audit-log events

What we don't have today — and why

  • HIPAA BAA — not a fit for this product right now; we don't recommend storing PHI in InventDB
  • FedRAMP — not on the roadmap
  • On-prem / self-hosted — not in Year 1; would slow product velocity, see pricing FAQ
  • BYOK for the AI model — not supported; we calibrate against Claude Opus 4.7 specifically
  • ISO 27001 — not on the immediate roadmap (most of the controls are in place; the audit isn't)

What happens if something goes wrong.

We're a small team and we're honest about what that means. We don't have a 24×7 NOC, we don't have a paid on-call rotation, and we don't run a follow-the-sun support model. What we do have: AWS-native alerting on the basic infrastructure signals, a single committed response window, and a written post-mortem for every confirmed incident that touched customer data.

Detection

Automated + reported

CloudWatch alarms on error rate, latency, and disk pressure page the team (KMS-access alerting joins the list once KMS integration ships). Customers can also email support@inventdb.com — treated the same as an alert.

Acknowledgement

Within 48 hours

Any customer-reported issue — critical or not — gets a human acknowledgement within 48 hours. That window is the same regardless of severity right now; we'd rather promise what we can keep than tier-segment a window we can't yet hold.

Notification

Within 72 hours of confirmation

Any incident with potential customer-data impact: email to your designated contact within 72 hours of confirmation, with what we know, what we're doing, and what you should do. This window is set by what the law expects (GDPR Article 33) and we hold ourselves to it.

A written post-mortem follows for every confirmed customer-data incident: timeline, root cause, blast radius, remediation, and the change we're making so it doesn't happen again. Customers get this in writing.

Coming as we grow the team: 24×7 on-call rotation, tier-segmented response windows (15 min for sev-1), and a public status page at status.inventdb.com. We'll publish a date when we can hold it, not before.

One commitment. No fine print.

What we commit to today

  • Email response: Any issue — critical or not, Personal or Business — gets a human acknowledgement within 48 hours. This is the only SLA we offer right now, and we offer it to every customer.
  • Backups: Daily encrypted snapshots; recovery rehearsed during routine deploys.
  • Notification on incidents: Within 72 hours of confirmation for anything that may have touched customer data (GDPR-aligned).
  • Channels: Email only today. A Business-customer Slack workspace is rolling out in the coming weeks. See the Contact section for the full list of inboxes.
We say one SLA and we mean it. No uptime guarantees, no service credits, no MSA exhibits with tier-by-tier response windows that a five-person team can't hold around the clock. Every customer — Personal $29 or Business $107+ — gets the same 48-hour response promise on every email.

Coming as we grow the team: a formal uptime SLA with service credits, tier-segmented response windows (1 hour for sev-1 production incidents on Business plans), and a published RPO / RTO. The infrastructure is built to support it (Multi-AZ on AWS, daily cross-region snapshots) — we just won't commit on paper to numbers we can't yet defend 24×7. As the team grows, this section will get more specific.

Security questions, vulnerability reports, takedowns.

Security contact: security@inventdb.com — acknowledged within 48 hours.

Vulnerability disclosure: We accept reports via email. Please include reproduction steps. We acknowledge within 48 hours, fix on a severity-driven timeline, and credit you in the release notes if you'd like.

Pen-test & assessment requests: Customers on Business plans can request our latest internal pen-test summary under NDA. Email security@inventdb.com.

Data-deletion request: Email privacy@inventdb.com from the registered owner email. Acknowledged within 48 hours; deletion completed within 30 days (the legal window).