# Itzal

> Enterprise data loss prevention for AI tool usage. Itzal detects sensitive
> data (PII, PHI, PCI, secrets) in text submitted to AI chat tools such as
> ChatGPT, Claude, Gemini and DeepSeek — entirely on the employee's device —
> and reports only a redacted event record. The raw content never crosses the
> network boundary.

Itzal runs as a browser extension (MV3) plus a local agent on macOS, Windows
and Linux, deployed via MDM. Detection is layered: deterministic patterns with
checksum validation, a quantized on-device NER model, and mapping to
compliance frameworks (HIPAA, GDPR/CCPA, PCI DSS, FERPA, GLBA, SOX, and
customer-defined internal rules). Policy actions are log, warn inline, or
block the submit before it is sent.

What is transmitted: finding type, count, confidence, a masked context
snippet, destination host, whether that destination is sanctioned, a risk
score, and the enforcement action that actually executed. What is never
transmitted: the prompt text, the model's response, attached file contents, or
the matched sensitive value in any reversible form.

## Core

- [How Itzal works](https://itzal.ai/product): capture, detect, redact, enforce — the end-to-end architecture
- [Privacy architecture](https://itzal.ai/privacy-architecture): exactly what crosses the network boundary and how it is enforced in CI
- [Coverage: what Itzal sees and doesn't](https://itzal.ai/product/coverage): published coverage matrix including uncovered surfaces — file uploads, pasted images, IDE assistants, mobile
- [Detection engine](https://itzal.ai/product/detection-engine): layers, published precision and recall targets per detector
- [Security](https://itzal.ai/security): mTLS with no HTTP fallback, encrypted local buffer, certificate-derived tenant identity

## Definitions

- [AI DLP](https://itzal.ai/glossary/ai-dlp): data loss prevention at the prompt layer
- [Shadow AI](https://itzal.ai/glossary/shadow-ai): AI tool use outside sanctioned channels
- [On-device redaction](https://itzal.ai/glossary/on-device-redaction): masking at the point of detection, irreversibly
- [Sanctioned AI tool](https://itzal.ai/glossary/sanctioned-ai-tool): the account-level distinction that should drive policy severity
- [Coverage gap](https://itzal.ai/glossary/coverage-gap): a surface a control does not inspect; silent ones cause incidents

## Answers

- [How to stop employees pasting sensitive data into ChatGPT](https://itzal.ai/answers/stop-employees-pasting-sensitive-data-into-chatgpt)
- [Does ChatGPT train on data employees paste into it?](https://itzal.ai/answers/does-chatgpt-train-on-pasted-company-data)
- [Is monitoring employee AI use legal?](https://itzal.ai/answers/is-monitoring-employee-ai-use-legal)
- [AI DLP vs CASB vs endpoint DLP](https://itzal.ai/answers/ai-dlp-vs-casb-vs-endpoint-dlp)
- [Can DLP work without reading employees' messages?](https://itzal.ai/answers/dlp-without-reading-employee-messages)

## Compliance

- [HIPAA and AI tools](https://itzal.ai/compliance/hipaa-and-ai-tools): what counts as PHI in a prompt; the BAA question
- [GDPR and AI tools](https://itzal.ai/compliance/gdpr-and-ai-tools): two distinct processing problems, lawful basis, DPIA, minimisation
- [PCI DSS and AI tools](https://itzal.ai/compliance/pci-dss-and-ai-tools): PAN and sensitive authentication data in prompts

## Optional

- [Deployment](https://itzal.ai/product/deployment): MDM packaging, staged rollout, legal prerequisites
- [Admin console](https://itzal.ai/product/admin-console): what security teams see, and SIEM integration
- [Evaluating AI DLP tools](https://itzal.ai/compare): the five questions that separate products, and how to verify a privacy claim yourself