Controls, evidence, and the gaps — verified 11 August 2026
Security
This page is written from an internal control audit dated 11 August 2026 in which every statement below was verified against a live system response, a configuration query, or a line of code — not from a marketing brief. Where a control is missing or not yet operating, it is listed in What we do not claim rather than omitted.
Encrypted in transit and at rest (AES). Never used for training. Never sold.
| PHI field encryption | Protected-health-information fields are individually encrypted with AES at the application layer before they reach the database, in addition to full-disk encryption of the database itself. |
|---|---|
| Key custody | The encryption key is held in AWS Secrets Manager and injected into the running task at start-up. It is not in source control, not in a container image, and not in an environment variable in any build artifact. |
| Database encryption | Amazon RDS storage encryption with a customer-managed KMS key, automatic annual key rotation enabled. Automated backups and snapshots inherit the same key. |
| Transport | TLS 1.2 and 1.3 only — TLS 1.0 and 1.1 are refused at the load balancer (ELBSecurityPolicy-TLS13-1-2-2021-06, live-tested). Plain HTTP is redirected to HTTPS. HSTS is sent with a one-year max-age including subdomains. |
| Database connections | Connections to the database require TLS (rds.force_ssl = 1). |
| Object storage | Both S3 buckets: KMS default encryption, versioning on, all four public-access blocks on. |
| Enforced by the database, not by queries | Tenant isolation is enforced by PostgreSQL Row-Level Security — a policy on the table itself, not a WHERE clause an engineer can forget. |
|---|---|
| The application role cannot bypass it | The serving database role is NOSUPERUSER and NOBYPASSRLS, and policies are declared FORCE ROW LEVEL SECURITY so they bind even the table owner. |
| Coverage | 113 of the 114 tenant-scoped tables carry an active, forced row-security policy. The one exception is listed under “What we do not claim” below. |
| Context cannot leak between requests | Tenant context is set as a transaction-local session variable and re-applied on every transaction begin, so a pooled connection cannot carry one tenant's context into another tenant's request. |
| Proven, not asserted | During the disaster-recovery drill on 2026-07-07, the application role without tenant context returned 0 rows from users, notes and episodes on a restored production snapshot. |
| Append-only audit trail | Every access and privileged action is written to an audit trail that is append-only, enforced by database triggers. The application role cannot modify or delete audit history. |
|---|---|
| No PHI in the audit trail | Audit records carry identifiers, action verbs and timestamps only — no clinical content — so they can be retained long term and exported to a customer SIEM. |
| Impersonation is stamped | Support “view as” sessions are read-only and every resulting audit row records the operator who initiated them. |
| Infrastructure audit | AWS CloudTrail is enabled across all regions with log-file validation and KMS encryption. Log groups retain 400 days. Alarms fire on root-account use, trail configuration change and administrative user creation. |
| Network position | Application servers run in private subnets with no public IP address. The database is not publicly accessible and accepts connections only from the application security group. |
|---|---|
| Identity | There are no long-lived IAM users in the production account. All human access is a short-lived assumed-role session. |
| Secrets | All ten runtime secrets are held in AWS Secrets Manager and injected at start-up. No secret value is stored in an environment variable, a container image, or source control. |
| Least privilege | The task role's model-inference permissions are scoped to named model families; its Cognito permission is scoped to one user-pool ARN and its email permission to one identity. The execution role can read exactly nine named secret ARNs. |
| Containers | The application container runs as a non-root user (uid 10001). |
| Availability | Database is Multi-AZ with deletion protection. Two application tasks run behind the load balancer with CPU target-tracking autoscaling. |
| SQL injection | All database access is parameterized. A repository-wide scan of all 211 raw-SQL call sites found zero dynamically constructed statements. |
|---|---|
| Code execution | The application contains no eval, exec, or shell-execution paths. |
| Static analysis | bandit runs on every change and gates at zero medium-or-higher findings. |
| Dependency, image and SBOM scanning | pip-audit, trivy and a CycloneDX SBOM run on every change to main and weekly on a schedule. Accepted advisories are documented individually with a reachability analysis and a named permanent fix. |
| Rate limiting | Per-IP rate limiting at the edge and in the application, with tighter limits on login, signup, token and password-reset endpoints. Request bodies are size-capped, including chunked uploads. |
| Response headers | X-Frame-Options: DENY, X-Content-Type-Options: nosniff, a restrictive Permissions-Policy (microphone on our own origin only; camera, geolocation, payment and USB denied), Referrer-Policy, and Cross-Origin-Opener-Policy. |
| Cross-site request forgery | Not applicable by design: the API authenticates with an Authorization: Bearer header, never an ambient cookie, so there is no cross-site request to forge. |
| Vulnerability disclosure | Published at /.well-known/security.txt (RFC 9116). Reports go to security@tebscribe.com. |
| Where inference runs | All large-language-model inference runs on Amazon Bedrock inside our own AWS account, under our signed AWS Business Associate Agreement. |
|---|---|
| Business Associate Agreements | Signed with AWS and with Deepgram, our speech-to-text provider (countersigned 2026-08-02). |
| No training on your data | Speech-to-text requests are sent with the vendor's model-improvement program explicitly opted out, on both the batch and the realtime path, guarded by an automated regression test that fails the build if the flag is dropped. |
| On-device option | Free-tier dictation runs on the device; that audio never leaves the phone. |
| Subprocessor register | We maintain a register naming every third party that processes customer data, what it processes and its BAA status. Available on request with a security questionnaire. |
| Backups | Automated database backups retained 35 days, with point-in-time recovery available. Backups inherit the database's customer-managed KMS key. |
|---|---|
| Restore has actually been tested | On 2026-07-07 we restored a production snapshot to a throwaway instance and measured a recovery time of 2 minutes 39 seconds against a four-hour objective; integrity was confirmed on record counts only. The procedure is written up as a repeatable runbook. This is one dated drill, not a demonstrated cadence. |
What we do not claim
A security review is faster when the gaps are on the first page. These are ours, as of 11 August 2026.
- Independent assurance. We hold no SOC 2 report of any type. We have engaged an external advisor and are working toward SOC 2 Type II; the observation window has not started. We will not imply a report exists before one does.
- Penetration testing. No third-party penetration test has been performed to date.
- Multi-factor authentication. Multi-factor authentication is available in our identity provider but is not currently required for every user. Requiring it is our next authentication change.
- Web application firewall. An AWS WAF is attached to the load balancer, but only the per-IP rate-limit rule takes action. The AWS managed rule groups run in count mode today — they log, they do not stop traffic. We will say so rather than list a firewall we are not enforcing.
- Threat detection. We do not run a security operations centre, a managed detection service, or around-the-clock human monitoring. Detection today is CloudTrail alarms on specific privileged events plus availability alarms.
- Content Security Policy. Our CSP is deployed in report-only mode. It is not enforcing, so we do not list it as a control.
- Audit-log integrity. The audit trail is append-only and hash-chained by design, but chain verification is not yet running on a schedule in production and the chain head is not yet anchored to write-once storage. Until both are true we claim append-only enforcement, and no more than that.
- Secret rotation. Automated rotation is not enabled on our runtime secrets. (One exception is permanent by design: rotating the field-encryption key would render existing encrypted records unreadable, so that key is re-keyed only through an envelope-encryption migration.)
- Row-level security coverage. One tenant-scoped table — a counter of clinical-calculator search terms, holding no note content — does not yet carry a row-security policy. It is the single gap in the wall and it is being closed.
- Recovery point objective. Our restore drill exercised the snapshot path. We have not yet drilled point-in-time recovery, so we do not publish a recovery-point objective.
- Enterprise single sign-on. SAML and OIDC enterprise identity providers are not available today. Google and Apple sign-in are.
- HIPAA. HIPAA has no certification scheme and no vendor can hold one. We operate as a business associate under signed Business Associate Agreements. We also make no Safe Harbor or Expert Determination claim about any data.
- End-to-end encryption. tebIQ does not provide end-to-end encryption and we will not describe it that way: the server decrypts clinical text in order to generate a note. Our separate on-device dictation product is the only place that claim applies.
Security questionnaires and reviews
Send questionnaires, BAA requests, architecture-review requests, and vendor-risk paperwork to security@tebscribe.com. We answer from the audit evidence behind this page and we will tell you when the honest answer is “not yet”.
Vulnerability reports: /.well-known/security.txt (RFC 9116) or the same address.
Controls verified 11 August 2026. This page is re-verified before it is changed.
Charts done before you leave the building.
tebIQ is physician-built AI documentation. Try it free — no demo call, no sales rep.
Start free