Use ChatGPT and Claude.
Keep the names inside.
Nymgate runs in your network and swaps names, IDs and secrets for tokens before a prompt leaves. Answers come back with the real values.
- Self-hosted on Kubernetes or OpenShift
- Offline install
- Zero phone-home
- Regulatory control mapping
Works with what you already run
The problem
Your controls say no. Your teams already use it.
Prompts carry customer data, so regulated teams block external AI. Staff paste it in anyway, or lose hours.
Nymgate gives security a yes it can defend.
How it works
Four steps on every request. If one fails, nothing leaves.
Find what is sensitive
Rules, a local model, your dictionaries. No LLM.
Swap it for a token
Same person, same token, per conversation.
Only tokens go out
Through your existing SASE, unchanged.
Read the real answer
Restored from your vault for the asker.
Try it yourself
Paste something sensitive.
Runs in your browser. Nothing is sent. Same validated rules as the gateway, plus your own. Names need the local NER model, so mark them here.
Select “Sarah Collins” and mark it: both mentions get the same token.
Your rules
The example catches the customer ID in the text.
| Token | Type | Value in your vault |
|---|
Two surfaces, one engine
Protect the chat window and the apps you build.
Browser extension
Managed on Edge and Chrome. Holds every ChatGPT and Claude send until it is tokenized.
Learn more →API gateway
Internal apps change one base URL. Streaming and tool calls included.
# before: https://api.openai.com/v1 client = OpenAI( base_url="https://nymgate.corp.internal/v1", api_key="pg_live_…", # gateway key ) client.responses.create(model="gpt-5", input=prompt)Learn more →
Compared
Three ways to handle external AI. One lets everyone say yes.
| Block AI tools | Cloud AI DLP | Nymgate | |
|---|---|---|---|
| Employees can use ChatGPT and Claude | ×No | ✓Yes | ✓Yes, in the native interface |
| Customer values stay away from the provider | ✓Yes, nothing is sent | ~Blocked or redacted, per policy | ✓Yes, replaced with tokens |
| Prompts are inspected inside your network | ~Nothing to inspect | ×Usually in the vendor's cloud | ✓Yes, on your servers |
| Answers come back with the real values | ~No answers | ×Redacted values are gone | ✓Restored for the person who asked |
| Your SASE stays as it is | ✓Yes | ~Often adds a proxy to the path | ✓Yes, it sits in front |
| When protection can't be applied | ✓Blocked | ~Depends on policy | ✓Fails closed |
“Cloud AI DLP” describes the typical SaaS inspection approach; individual products differ.
Architecture
Everything that can reverse a token stays in your network.
Remote staff connect through your ZTNA or VPN. No public endpoint, no vendor tunnel.
What the architecture guarantees
- Every protected request passes the gateway in your environment
- Detected values are transformed before transmission
- Mappings and keys never leave your network
- Unsupported or unhealthy flows fail closed
What we measure and publish
- Precision and recall per entity type
- Results on a Turkish/English corpus with the model version
- A local harness so you can score it on your own data
What we will not claim
- That no personal data can ever leak
- That tokenized data is anonymous under data protection law
- That we replace your legal basis for processing
Paid pilot
From install to a decision in six to eight weeks.
Install
Helm or Docker Compose, offline images in your registry, checks for DNS, TLS, OIDC and egress.
Browser cohort
50–200 managed users on ChatGPT with tokenize and restore.
One internal app
Switch its endpoint. Score detection on your own labelled sample.
Review
Against success criteria agreed on day one. Then an annual licence, or a clean exit.
Apply for the pilot
A few design partners at a time, each with an AI workflow their controls block today.
Questions security teams ask
FAQ
Does any of our data reach Nymgate?
No. The gateway, detector and vault run in your Kubernetes or OpenShift cluster. Licences are signed offline and there is no telemetry.
What happens if the gateway is down?
Protected sends stop and the text stays in the composer. Nothing is released to the provider untransformed.
Do we have to replace our SASE or DLP?
No. Nymgate sits in front of it. FortiSASE, Zscaler or Netskope keep enforcing which AI tools managed and unmanaged devices may reach.
Are file uploads protected?
Not in the first release. Upload controls are disabled on managed surfaces so nobody assumes a file is protected. Text is covered today.
How long can a token be reversed?
For as long as its mapping lives in your vault: 24 hours for stateless API calls and 7 days for chats by default, refreshed on use. After that it shows as an expired placeholder.