/ Security
Security practices written as practices, not as badges.
- Last reviewed
- September 11, 2026
- Scope
- novireli.com and engagements delivered by Novireli
- Report a vulnerability
- support@novireli.com
[ 01 ]Principles
Least access, granted late
We ask for the narrowest access that completes the work, prefer sandbox and test credentials, and request production access only when a task genuinely requires it.
Your data is not training data
We do not use client data to train models, and we configure model providers to use zero-retention or no-training options where those options exist for the chosen provider.
Secrets stay in secret managers
Credentials live in the platform's secret storage or your own vault. They are not committed to repositories, pasted into chat tools, or sent by email.
Humans stay in the loop where it matters
Systems that affect money, contracts, customers or compliance are designed with an approval step and a confidence threshold, not left to decide alone.
Auditable by design
Model calls, inputs and decisions are logged so that any output can be traced to what produced it. Retention windows are agreed with you rather than assumed.
Clean exit
Access is revoked at the end of an engagement, working copies of your data are deleted on request, and you keep the code and documentation.
[ 02 ]During an engagement
What happens to your access and your data.
- Access requests
- We ask for the narrowest role that completes the task, prefer sandbox or staging environments, and use your identity provider and access-review process rather than shared logins. Production access is requested per task and given up when the task is done.
- Credentials
- Secrets are stored in the hosting platform's secret manager or in your own vault. They are never committed to a repository, pasted into a chat tool, or sent over email. If a secret is ever exposed, we tell you and rotate it immediately.
- Devices
- Work happens on full-disk-encrypted machines with automatic screen lock, current operating-system updates, and multi-factor authentication on every account used for client work.
- Client data copies
- We prefer synthetic or redacted samples. Where real data is needed for evaluation, it is the minimum required, kept in a dedicated project location, and deleted at the end of the engagement or sooner on request.
- Subprocessors
- The third parties involved in a project — hosting, model providers, error reporting — are listed in the proposal before work starts. We do not introduce a new subprocessor into your system without telling you.
- Contractors
- If a specialist is brought in, they are bound by the same confidentiality terms we owe you, and you are told before they are given access.
- Offboarding
- At the end of an engagement, access is revoked, working copies are deleted on request, and you receive confirmation in writing.
[ 03 ]AI-specific safeguards
The risks that come with the model.
- Your content is not used to train models. We select providers and settings that exclude API traffic from training, and configure zero or minimal retention where the provider offers it.
- Prompts are treated as untrusted input. Content retrieved from documents, email or the web is never allowed to trigger privileged actions on its own.
- Confidence thresholds route uncertain cases to a person rather than letting the system guess, and anything touching money, contracts or customers has an approval step.
- Permissions are enforced at retrieval time, so an answer engine cannot surface a document the person asking is not allowed to read.
- Model calls, inputs, outputs and costs are logged so that any result can be traced back to what produced it. Retention windows for those logs are agreed with you.
- Evaluation sets are built from your real cases and re-run when prompts, models or dependencies change, so quality regressions are caught rather than discovered by a customer.
A limit worth stating
No control makes a language model deterministic. We reduce the blast radius of being wrong; we do not claim it will not happen, and we design the surrounding system on that assumption.
[ 04 ]This website
The smallest attack surface we could reasonably ship.
- Static by design
- novireli.com is a statically rendered marketing site. There are no user accounts, no login, no database and no file uploads.
- No forms
- Nothing on this site collects input. Every contact route is a mailto link that opens your own mail client.
- No cookies or analytics
- We set no cookies, run no analytics, and load no advertising or social pixels. There are no third-party scripts on any page.
- No external assets
- All imagery is inline SVG and CSS. Fonts are self-hosted from our own domain at build time, so no request leaves the site to a third-party asset host.
- Transport security
- The site is served over HTTPS by our hosting provider, which terminates TLS and redirects plain HTTP.
- Dependencies
- Framework and dependency updates are applied on a regular cadence, and security advisory alerts are reviewed when they are raised.
[ 05 ]Vulnerability disclosure
Found something? Tell us and we will fix it.
What we ask of you
- Give us a way to reproduce the issue: URL, request, and what you expected to happen instead
- Test only against your own accounts and data, and stop as soon as you have proved the issue exists
- Do not run denial-of-service tests, spam, social engineering, or physical attacks
- Give us a reasonable window to fix the issue before publishing details
What you get from us
- A written acknowledgement after we review your report
- An assessment and remediation plan based on the issue's severity and scope
- Credit in the fix notes if you want it, and no legal action for good-faith research that follows the guidelines above
If something goes wrong on our side
If we become aware of a security incident affecting a client's data or a system we maintain, we will notify the affected client without undue delay, and within any shorter period set out in the signed agreement. Notification includes what we know, what we do not yet know, the immediate steps taken, and the plan to close it out.
We would rather send an early notification with incomplete information than a tidy one a week later.
[ 06 ]What we do not claim
The certifications we do not hold.
- SOC 2 Type I or Type II attestation
- ISO/IEC 27001 certification
- HIPAA, PCI DSS or FedRAMP compliance
- A third-party penetration test of this website
- A 24/7 staffed security operations centre
- Any uptime or availability percentage
If your procurement process requires an audited framework, we will tell you at the first call that we cannot meet that requirement today rather than starting a process that fails at legal review. This page will be updated if that ever changes.