// ZDZAISHIELD · LOCAL-AI DLP FOR SHADOW AI

Your team is pasting
sensitive data into ChatGPT
right now.

We block it. ZDZAIShield is a DLP with local AI that inspects every prompt and stops CPF and CNPJ numbers, cards, passwords, and secrets from reaching public AIs — with an HTTP 403, before the data leaves your network.

7 public AI services covered 100% local AI · Ollama Data never leaves the network LGPD by design
7
public AI services monitored
403
HTTP on sensitive data submissions
100%
local inference, no cloud
~98%
savings vs. equivalent cloud
// The invisible problem

Shadow AI: the leak nobody sees happening

It is not an attack. It is routine. Every day, well-meaning people paste customer data into a public AI to "speed things up" — and the data leaves your network without passing through any control.

📋

CPF and CNPJ in the prompt

"Summarize this record for me." A CPF, a CNPJ, an entire contract pasted into a text box that sends everything to third-party servers, outside the country.

🔑

Secrets and credentials

"Why doesn't this code run?" — and along go API keys, passwords, and connection strings. What was a bug becomes corporate secret exposure.

🩺

Sensitive and health data

Medical records, reports, financial and legal data. Categories the LGPD treats as sensitive (art. 11) leaving without a legal basis, without a trail, without consent.

⚠️

The blind spot is not the AI. It is the lack of control at the exit.

Blocking public AI at the firewall does not work: people use their phone, their personal account, the next tool. What is missing is inspecting the content of what goes out — and deciding, in real time, what is allowed and what is not.

// How it works

A DLP that understands the text — and blocks with a 403

ZDZAIShield sits in the path between your network and the public AIs. It inspects each prompt in two layers and, if sensitive data is found, returns an HTTP 403 — the request never reaches the external provider.

01 · Intercept

On the egress path

Every prompt bound for a public AI goes through the proxy before leaving the network. ChatGPT, Gemini, Claude, Copilot, DeepSeek, Perplexity, HuggingFace — 7 services covered.

02 · Inspect

Regex + local AI

Known patterns (CPF, CNPJ, cards, keys) are caught by regex. Whatever slips through goes to NLP sensitivity classification on a local LLM.

03 · Decide

Sensitive? 403.

If the content is classified as sensitive, the proxy blocks it with an HTTP 403 and explains why. The data does not leave. Clean prompts go through normally.

04 · Log

A trail for the DPO

Every block becomes an auditable event: what was stopped, to which service, when. Governance to prove compliance, not just promise it.

NLP classification runs on a local LLM via Ollama — the inspection of your data is never outsourced to the cloud.

// Blocking demo

A CPF trying to reach ChatGPT — stopped

See what happens when someone pastes personal data into a prompt bound for a public AI.

POST chat.openai.com/prompt · via ZDZAIShield
employee "Analyze this customer record: name John Doe, CPF 000.000.000-00, and write a summary of their history."
aishield inspecting content → regex: match · local NLP classification: PERSONAL DATA (high confidence)
HTTP 403 · FORBIDDEN — submission blocked
CPF / personal data detected in a prompt bound for a public AI (ChatGPT). The request was stopped and never left the network. Event logged for the audit trail. Remove the sensitive data or use the approved local AI.

(illustrative example · fictitious, masked CPF · values do not represent real customer data)

// Why local

The AI that inspects your data lives in your infrastructure

A DLP that ships your content to the cloud for "analysis" recreates the very problem it is supposed to solve. ZDZAIShield runs the classification LLM inside your network, with Ollama. Private inference, isolated per client.

  • Data sovereignty. Inspected content never leaves the network — not to block, not to classify.
  • Inference isolated per client. Your own local LLM instance, with no third-party multitenancy in between.
  • On-premises-first. Runs in your mini data center or dedicated server, with no dependence on a cloud bill.
  • "Local Ollama — data never leaves the network." It is the design rule, not an option.
~98%
savings vs. equivalent cloud
The same on-premises foundation that runs ZDZAIShield delivers local LLM inference at a fraction of cloud cost — without giving up sovereignty.
// LGPD & governance

From policy to auditable event

Compliance is not a PDF in a drawer. ZDZAIShield turns your AI usage policy into a technical control that acts on its own — and leaves a trail.

📐

Policy that becomes control

The rule of "never paste customer data into a public AI" stops depending on training and goodwill: it is enforced in real time, on the way out.

🧾

Audit trail

Every attempt and every block is logged: which data category, to which service, when. The DPO gets evidence of compliance, not assumptions.

🩺

Sensitive data (art. 11)

Directly addresses the processing of sensitive categories — including health data — preventing them from leaving the network without a legal basis.

// Frequently asked questions

What CISOs and DPOs ask

Which public AIs does ZDZAIShield cover? +
Seven public AI services: ChatGPT, Gemini, Claude, Copilot, DeepSeek, Perplexity, and HuggingFace. Any prompt bound for any of them goes through inspection before leaving the network.
Is my data sent anywhere to be analyzed? +
No. Sensitivity classification runs on a local LLM via Ollama, inside your infrastructure. Inspected content never leaves the network — not to block, not to classify. Inference is private and can be isolated per client.
How does the blocking work in practice? +
ZDZAIShield sits on the egress path. It inspects each prompt in two layers — regex for known patterns (CPF, CNPJ, cards, keys) and local-AI NLP classification for what slips through. If the content is sensitive, the request gets an HTTP 403 and never reaches the external provider. Prompts without sensitive data go through normally.
Does this help me with the LGPD? +
Yes. It enforces your AI usage policy as a technical control and generates an audit trail for every block — the kind of evidence a DPO needs. It directly addresses the processing of personal and sensitive data, including health data (art. 11), preventing it from leaving without a legal basis.
Why run the AI locally instead of in the cloud? +
For sovereignty and for cost. A DLP that sends your content to the cloud recreates the problem it is supposed to solve. Running locally, the data never leaves — and the same on-premises foundation delivers LLM inference at roughly 98% savings versus equivalent cloud.
// Shadow AI risk assessment

Find out what is already leaving your network

A risk assessment to map where your team uses public AI, what kind of data is exposed, and how ZDZAIShield closes that door. AI + People: the machine blocks, your governance decides.