Security Pack
Prepared for security review. It covers where SecuriX runs, what data it touches, how long that data is kept, who else is involved, and what happens when something goes wrong.
Last reviewed: 2026-08-26
What We Hold, and What We Don't
Stated plainly and up front. SecuriX is an early-stage company and does not hold SOC 2 or ISO certification today. Where this site references those frameworks, it means the controls are designed against them — not that an audit has been completed.
Current status
Controls are designed against SOC 2 criteria. No audit has been completed and no report exists.
Required for Google restricted OAuth scopes. Assessment underway; not yet complete.
No certification. Referenced elsewhere on this site only as a framework we design against.
AI management system standard. Under evaluation as a roadmap item.
We do not yet have a counsel-reviewed Data Processing Agreement. One is being prepared. In self-hosted and VPC deployments SecuriX processes no customer personal data, which narrows the scope considerably.
Where Requests Go
Two enforcement points. The LLM Gateway governs traffic to model providers; the Secure MCP Server governs what agents may do against your own systems. Both evaluate policy before data moves and again on the way back.
- —DLP policy engine — PII detected and redacted
- —Credential substitution — provider keys vaulted
- —Model whitelisting per team and role
- —Request / response streaming and logging
- —OPA / Rego policy evaluated pre- and post-call
- —Agent over-reach prevention (e.g. recursive delete)
- —Tool-call conversion to provider REST APIs
- —Response redaction before the model sees data
It Runs Where You Put It
Your data centre. No SecuriX-operated subprocessor sits in the data path, and no prompt content leaves your network.
Your cloud tenancy, your key management, your network policy. Residency follows the region you deploy into.
Managed by us for evaluation. The subprocessor list below applies only to this model.
What We Touch and Keep
Data handling
User identity, team, model, token counts, cost, policy decision, timestamp, and a cryptographic hash of the prompt. This is what powers cost attribution, budgets, and the audit trail.
Traffic that passes policy cleanly has its metadata recorded but its content is not stored. You retain the material a compliance review needs without warehousing every employee conversation. In self-hosted and VPC deployments that store sits inside your own network.
Retention is set by the customer to match their own records policy. In self-hosted and VPC deployments the data never leaves your infrastructure, so deletion is under your control.
DLP runs on the request before it leaves the gateway, and on tool responses before the model sees them.
Data keys are wrapped by a KMS-held master key. Plaintext keys are never persisted.
SecuriX operates no models of its own. Provider-side training terms are governed by the customer's own contract with OpenAI, Anthropic, or Azure.
Applies to the managed offering only. Self-hosted and VPC deployments reside wherever you deploy them, so residency requirements are met by your own choice of region.
Encryption & key management
Configured via GCP_KMS_KEY_PATH.
Customer-operated Vault — the master key never leaves your control.
All traffic in transit, including calls onward to model providers.
Who Else Is Involved
These apply to SecuriX Cloud only. In self-hosted and VPC deployments there is no SecuriX-operated subprocessor in the data path. Model providers you connect are contracted directly by you, not by us.
| Subprocessor | Purpose | Data touched |
|---|---|---|
Google Cloud Platform | Application hosting, managed database, key management, messaging | Application data, encrypted credentials, audit records |
PostHog (US) | Product analytics on the marketing site and dashboard | Usage telemetry and page events. No prompt content. |
Neon (US region) | Managed PostgreSQL | Application data and audit records for flagged traffic |
Formspree | Demo and contact form submissions on securix.app | Name, email, and message submitted through the marketing site. No product data. |
- Self-hosted deployments bundle PostgreSQL — Neon is not involved.
- SecuriX sends no transactional email from the product. Formspree applies only to the marketing website.
- Model providers you connect are contracted directly by you and are not SecuriX subprocessors.
When Something Goes Wrong
Response commitments
Reports are triaged during Indian Standard Time business hours. Report anything you believe to be a vulnerability — we will not pursue researchers acting in good faith.
Ahead of the 72 hours required under GDPR Art. 33. Notification goes to your nominated security contact.
Sent by email once the immediate incident is contained, covering what happened, what data was involved, and what changed as a result.
Severity model
Confirmed unauthorised access to customer data, or a full outage of the gateway in production.
Triage begins immediately on detection. Customer notified within 48 hours of confirmation.
Vulnerability that could lead to data exposure, or degraded policy enforcement — for example DLP failing open.
Triage within one business day. Fix prioritised above all feature work.
Security defect with no immediate exposure — a dependency advisory, or a hardening gap.
Triage within three business days. Scheduled into the next release cycle.
Informational finding or recommended improvement.
Acknowledged and logged in the security backlog.
Available for Review
Not yet available. We will sign your standard DPA where its terms are ones we can meet, and we will tell you plainly where they are not.
Completed from this pack for your review. We will also complete your own questionnaire rather than asking you to accept ours.
Questions From Your Security Team
We will complete your own questionnaire rather than asking you to accept ours, and we will answer “not yet” where that is the honest answer.
Talk to Us