What stays and what leaves

Every request makes the same round trip: encoded in your environment, reasoned on by the external model you choose, decoded back in your environment.

How sensitive values are found

Sensitive values are found locally, in one or more ways: flaxmor’s advanced detection, custom rules for your company, or your existing detection system. You can also add any entity by hand. Everything found or added is encoded before the prompt or file leaves.

How the encoding works

Each protected value is transformed by a mathematical algorithm that includes randomization, combined with AI engineering, so the model can still do the work on the encoded data. Not cryptography.

Where it runs

On-prem or in your cloud: your choice. Your team uses flaxmor through a custom interface or an API.

Requirements: a 16-core CPU with 32 GB of RAM serves more than 200 users.

Network: only encoded data leaves your environment. It goes through flaxmor’s routing service to the model provider you choose.

Your model provider

flaxmor works with OpenAI, Anthropic Claude, Google Gemini, Microsoft Copilot, Mistral and more. You connect it with your own API key for the provider.

What flaxmor can access

Encoded prompts pass through flaxmor’s routing service on their way to your model provider. That is the only thing that reaches us.

The decode key stays only in your environment, so flaxmor, the company, never sees your real data.

Answer quality and how to check it

We compare the same prompt and file sent through the model’s own interface and through flaxmor. Across tens of thousands of checks, the protected answer scored the same or better than the model’s own interface.

Answers are graded by an LLM acting as judge, which scores each answer out of ten on how well it did the job.

Run the check yourself

1. Check the provider’s logs. Send a mock prompt and file, then open the model’s API logs. Confirm the values that arrived can’t be worked back to the originals.
2. Run it again, unprotected. Send the same prompt and file to the same model through it's online interface.
3. Compare both answers. Have the model score each answer out of ten on how well they did compared to each other.

Use mock files with the same structure for the unprotected run, never your real data.

Compliance status

SOC 2 Type II

In progress, expected by the end of November 2026

ISO 27001

In progress, expected by the end of November 2026

This website

This page is about the flaxmor product. For what happens to information you give this website, read the Privacy Policy. Our contact form delivers submissions to our inbox through Google Workspace, our email provider.

Security questions

Email contact@flaxmor.com, or talk to us about a design partnership and bring your security team.