India
Aadhaar and PAN, with checksum validation to confirm a structurally valid identifier before flagging.
AI Data Loss Prevention
Stop sensitive data from leaking through your AI.
AI-DLP for LLM and agentic apps: detect and redact PII, secrets, and regulated identifiers in prompts, responses, and tool calls, all inside your own boundary, before anything reaches a model vendor.
znyx · ai data loss preventionin boundary
PII · 65+ types
ZNYXRedacted in boundary
Definition
What is AI data loss prevention?
AI data loss prevention is about stopping sensitive data from leaving your organization through AI surfaces: the prompts users send, the responses models return, and the tool calls agents make. ZNYX detects PII, secrets, and regulated identifiers on every one of those paths and redacts them in your own environment, so the raw data never has to be handed to an external vendor.
PII detection & redaction
65+ PII types, including checksum-validated regional IDs
The PII detector recognizes 65+ sensitive data types and validates regional identifiers against their check-digit algorithms before flagging, so you catch real identifiers without drowning in false positives.
United Kingdom
National Insurance Number (NINO) and NHS number, recognized by format and check digits.
Brazil
CPF, validated against its check-digit algorithm to cut false positives on look-alike numbers.
Contact data
Emails, phone numbers, and postal addresses across regions and formats.
Financial
Payment card numbers, bank and routing identifiers, and other monetary references.
Network & device
IP addresses and other technical identifiers that can re-identify an individual.
Partial masking
Preserve enough structure for the workflow while hiding the sensitive value - useful when an agent needs to know a field is present without seeing the underlying data.
Full redaction
Replace the matched value entirely with a placeholder so nothing recoverable remains in the prompt, response, or tool payload.
Secrets & credentials
Block secrets and credentials in prompts and responses
A dedicated secrets and exfiltration detector catches credential-shaped strings on both sides of the request: a leaked key pasted into a prompt, or a credential a model surfaces in its output.
- Flags API keys, tokens, private keys, and other credential patterns in inputs and outputs.
- Choose the remediation per policy: block, redact, mask, re-ask, refrain, or ask a human.
- Runs through the same deterministic to ML to LLM-judge ladder as every other detector.
- Pairs with the output egress gate so a surfaced secret is caught before it leaves.
Output egress / DLP gate
An egress gate with a host allowlist and audit trail
Detection catches sensitive content; the egress gate decides whether a payload is allowed to leave. Together they form a DLP control over outbound model responses and tool calls.
Host allowlist
The egress gate inspects outbound payloads and enforces an allowlist of permitted destinations, so data only flows to hosts you have explicitly approved.
Audited decisions
Every block, redaction, and pass is captured as an audit event with the trace id and policy decision, building a defensible record of what was inspected and why.
Fail-closed gate
The egress and scorecard gate is fail-closed: when a check cannot complete, the safe outcome is to withhold the payload rather than let unverified data leave.
In-boundary redaction
Redact in your boundary, data never goes to a vendor
The most important property of AI-DLP is where the work happens. ZNYX runs detection and redaction inside the self-hosted runtime, and an in-boundary inference sidecar keeps ML and judge models in your VPC. Sensitive values are removed before any payload is forwarded, so you are not relying on a third party to see the raw data in order to scrub it.
GDPR alignment: built for data minimization
Forward only what is needed
Redacting PII before a prompt leaves means you process and transmit only the data the task requires, the core of GDPR data minimization.
Evidence by default
Every detection and egress decision is recorded, so you can show the control was applied without exposing the underlying values.
Air-gappable
For the strictest environments, self-host the control plane too and run fully private, even air-gapped.
FAQ
AI-DLP frequently asked questions
Quick answers on PII redaction, secrets blocking, the egress gate, where redaction runs, and how AI-DLP supports GDPR data minimization.
AI-DLP is the practice of stopping sensitive data - PII, secrets, credentials, and regulated identifiers - from leaving your organization through LLM prompts, model responses, and the tool calls agents make. ZNYX inspects every request path, detects and redacts sensitive content in your own boundary, and enforces an output egress gate before any payload reaches an external destination.
The detection and redaction step runs inside the self-hosted runtime, so prompts are scanned and rewritten before they are forwarded anywhere. You choose partial masking (preserve format, hide the value) or full redaction per detector. Because the work happens in your environment, the raw PII never has to be sent to a third-party vendor for it to be removed.
The PII detector covers 65+ types, including checksum-validated regional IDs such as India Aadhaar and PAN, UK NINO and NHS numbers, and Brazil CPF, alongside emails, phone numbers, payment card numbers, and IP addresses. Checksum validation reduces false positives by confirming an identifier is structurally valid before it is flagged.
Yes. A dedicated secrets and exfiltration detector flags API keys, tokens, private keys, and credential-shaped strings in both inputs and outputs. You can configure it to block, redact, or re-ask, so a leaked key in a prompt or a credential surfaced in a response is stopped before it leaves the boundary.
Redacting PII before it is sent to a model means you process and forward only the data the task actually needs, which directly supports the GDPR data-minimization principle. Every detection and egress decision is recorded as an audit event, giving you the evidence trail to demonstrate the control was applied.