NymGateCustomer-hosted AI privacy enforcement
Use AI.
Keep sensitive data inside your environment.
NymGate detects and transforms sensitive data before it reaches ChatGPT, Claude, coding agents or model APIs, then restores it for authorised users. Customer-hosted, policy-controlled, and built to work with your existing security stack.
- Customer-hosted: on-prem, VPC, Kubernetes, OpenShift
- Browser AI, coding agents and model APIs
- Block, allow + audit, or observe
- No SaaS control plane
Works with what you already run
The problem, and why now
Your teams already use AI. Blocking it isn't a plan.
Data leaves in prompts
Analysts and support staff paste customer records into chat because it is the fastest way to get work done.
Agents read everything
Coding agents open files, logs and configs, and send what they read to the model, credentials included.
Blocking moves the risk
A block list pushes AI to personal devices and accounts, where you have no control and no record.
The platform
Observe. Transform. Enforce. Prove.
One control layer across browser AI, coding agents and model APIs, run and managed inside your environment.
Three surfaces, one engine
Not a browser plug-in. A control layer for every way AI is used.
Browser AI
Managed Chrome and Edge extension with browser-native enforcement.
Learn more →Coding agents
Prompts, tool calls and tool results, centrally configured.
Learn more →AI APIs
Provider-native endpoints. JSON, streaming and tool calls.
# 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 →
Transform, don't just block
Keep the model useful. Keep the values inside.
Why did the payment from Sarah Collins's account GB82 WEST 1234 5698 7654 32 fail?
Why did the payment from [REDACTED]'s account [REDACTED] fail?
Why did the payment from ⟦PERSON:7F31:0001⟧'s account ⟦IBAN:7F31:0002⟧ fail?
Sarah Collins's transfer from GB82 WEST 1234 5698 7654 32 was rejected: the beneficiary details did not match.
- Same value, same token, for the whole conversation
- Deterministic and reversible, only inside your vault
- Restored only for the authorised user
- Mappings expire on a schedule you set
- Stored provider chats hold tokens, not originals
Security enforcement
Real enforcement behind the transformation.
For supported and validated integrations, NymGate verifies that protected outbound AI traffic matches an approved transformed request before transmission.
- 01AI pagechatgpt.com
- 02Transport transformationrequest recognised
- 03NymGatedetect · transform
- 04One-shot approvalexact body, single use
- 05Browser enforcementcompare, then release
- 06Providertransformed values only
Anything that does not match an approval is cancelled before it is sent.
Block, allow with audit, or observe.
Block
Strict fail-closed. Nothing leaves unless it was transformed and approved.
For High-assurance and regulated workflowsAllow + Audit
Work continues when enforcement cannot complete, and an explicit security event is raised.
For Business continuity with accountabilityObserve
Measure AI usage and policy impact before you enforce anything.
For Discovery and rollout planningPolicy per failure class
# nymgate policy · v42 enforcement: default: block on_failure: detector_unavailable: block gateway_unavailable: block unknown_schema: allow_with_audit unverified_transport: block groups: research-pilot: observe
Fail-open behaviour is always explicit policy, never a silent fallback.
Customer-hosted by design
Your data plane stays yours.
NymGate does not require prompts, mappings or detection workloads to be sent to a NymGate SaaS control plane.
What the architecture guarantees
- Protected traffic passes the gateway in your environment
- Detected values are transformed before transmission
- Mappings and keys never leave your network
- Enforcement failures follow explicit policy
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
- 100% leak prevention
- That tokenized data is anonymous under data protection law
- That we replace your legal basis for processing
Why security teams choose NymGate
From banned to governed.
Say yes without an exception
Approve external AI with a control you can defend, not a policy waiver.
One policy, three surfaces
Browser AI, coding agents and APIs under the same rules and the same audit.
Keep the model useful
Tokens keep who and which account intact, so answers stay specific.
Answer the auditor
Show every decision with metadata, never with transcripts.
Deploy where SaaS is not allowed
Air-gapped, on-premises or VPC, with no vendor in the data path.
Keep what already works
Your SASE, DLP, EDR and MDM stay exactly where they are.
Use cases
Concrete workflows, across the company.
Coding agents without leaked secrets
Secrets, customer IDs and operational data stay out of agent context.
Analysis on real cases
Analysts use external AI without sending account or customer identifiers.
Investigate tickets faster
Support teams work cases with AI; identifiers are transformed first.
Draft with personal data protected
Personal and organisation-specific values never reach the provider.
Integrations
Add AI-specific controls without replacing what already works.
- ChatGPT
- Claude
- Claude Code
- Codex
- Cursor
- OpenAI
- Anthropic
- Microsoft Entra ID
- Chrome Enterprise
- Microsoft Edge
- Kubernetes
- OpenShift
- Redis
- Your HTTP / SASE proxy
Compared
Where NymGate fits next to what you may already have.
| Traditional DLP | Generic LLM gateway | NymGate | |
|---|---|---|---|
| Browser AI (ChatGPT, Claude) | ~Typically block or redact | ×Typically API traffic only | ✓Managed extension, enforced |
| Coding agents | ~Varies | ~When routed through it | ✓Claude Code, Codex, Cursor |
| Direct model APIs | ×Typically not in the path | ✓Yes | ✓Provider-native |
| Reversible, conversation-consistent transformation | ×Typically block or redact | ~Varies | ✓Scoped tokens, restored for the user |
| Streaming, tool calls and tool results | ×Not designed for it | ~Varies | ✓Restored mid-stream |
| Browser enforcement of the outbound request | ~Endpoint or proxy based | ×Not applicable | ✓Exact-body, one-shot approval |
| Detection hosted by the customer | ~Varies | ~Varies | ✓In your environment, CPU-only |
| Turkish sensitive data | ~Varies | ~Varies | ✓TCKN, VKN, IBAN, Turkish-first NER |
| Coexists with your SASE | ✓Often part of it | ~Varies | ✓Alongside it, unchanged |
| Metadata-only audit | ~Varies | ~Varies | ✓No prompts, responses or mappings |
Columns describe typical approaches in each category; individual products differ.
Deployment
Built for environments where “send it to our SaaS” is not an acceptable architecture.
- OCI images
- Helm
- Non-root
- CPU-only inference
- Redis
- Readiness and liveness
- Corporate HTTPS proxy
- Custom CA bundle
- Signed versions
- Docker for pilots
Compliance, stated carefully
Reduce the amount of sensitive data exposed to external AI services and support your organisation's data-minimisation and AI-governance controls.
NymGate does not make you KVKK compliant on its own, does not make an international transfer lawful, and does not anonymise all personal data.
Resources
For the security review.
Demo and enterprise pilot
See it on your workflows. Decide in six to eight weeks.
Install
Helm on Kubernetes or OpenShift, offline images in your registry, Entra ID connected.
Observe
Browser cohort in observe mode. See real AI usage and policy impact.
Enforce
Block or allow + audit for browsers, one coding team and one internal app.
Review
Against success criteria agreed on day one. Then an annual licence, or a clean exit.
Book a demo
A 30-minute walkthrough on your use cases with an engineer, then a pilot plan if it fits. Pick a time on the next screen.
Questions security teams ask
FAQ
Does any of our data reach NymGate the company?
No. Gateway, detection, vault, policy and audit run in your environment. There is no SaaS control plane and no telemetry.
What happens if the gateway or detector is unavailable?
Your policy decides, per failure class: block (the default), allow with audit, or observe. It is never a silent fallback.
Is this another DLP? Do we replace our SASE?
No. Your web gateway, SASE, EDR, MDM and DLP stay in place. NymGate adds the AI-specific layer: conversations, streams, tool calls and coding-agent traffic.
Does NymGate make us KVKK compliant?
No product does on its own. NymGate reduces the sensitive data that reaches external AI and gives you customer-hosted processing, retention control and audit to support your assessment.
Which coding agents are supported?
Claude Code, Codex and Cursor, configured centrally with API-key profiles. Prompts, tool-call arguments and tool results are transformed.
Are file uploads protected?
Not yet. Images and files are rejected or disabled on protected surfaces so nobody assumes they were transformed. Text is covered.
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.
Don't block AI. Control what leaves.
Give employees and developers the AI tools they already use, without treating every prompt as safe by default.