Data Loss Prevention

Detect sensitive data
before it leaves

Inspect supported prompts, responses and uploads on your infrastructure, then allow, warn, block or redact according to policy.

Request an AI Security Assessment ← Back to SaaS AI
7
ML detection engines, all on-premises
40+
PII entity types detected
<1ms
Regex detection latency
~35ms
Average ML inference latency
Detection Engines

Seven purpose-built models.
All running on your hardware.

Each engine is optimized for a different threat class. Together they cover the full spectrum — from a leaked SSN to a prompt injection buried in a 2,000-token system prompt.

Engine 1 — Transformer NLU

Prompt Injection & Jailbreaks

Fine-tuned transformer model detects adversarial instructions embedded in prompts — override directives, roleplay jailbreaks, indirect injection via retrieved content, and multi-turn manipulation sequences.

Injection
Engine 2 — Content Classifier

Toxicity & Threats

Fine-tuned content classifier identifies toxic language, harassment, hate speech, explicit threats, and self-harm content in both inbound prompts and outbound AI responses — in either direction.

Toxicity
Engine 3 — Entity Recognition (NER)

40+ PII Entity Types

Named Entity Recognition model trained on enterprise data patterns. Detects: names, email addresses, phone numbers, SSN, passport numbers, national IDs, credit card numbers, bank accounts, crypto wallet addresses, passwords, OAuth tokens, and 30+ additional entity types.

PII · 40+ types
Engine 4 — URL Classifier

Malicious URL Detection

Deep learning classifier trained on malicious URL patterns — phishing domains, command-and-control endpoints, URL-encoded payloads, and obfuscated redirects embedded in prompts or responses.

Malicious URLs
Engine 5 — Semantic Classifier

Custom Semantic Topics

Zero-shot topic classifier (INT8-quantized for low latency). Describe a prohibited topic in plain English and the model detects it without retraining. Comes with 9 built-in topics and supports unlimited custom additions.

Zero-shot
Engine 6 — Regex

Secrets & Credentials

High-precision pattern library covering: GitHub personal access tokens, AWS access keys and secrets, GCP service account keys, Azure credentials, JWT tokens, generic API key patterns, SSH private keys, and source code fingerprints. Detects in under 1ms.

Secrets
Engine 7 — Image OCR

Image OCR on Multimodal Prompts

Screenshots, scanned documents, and images pasted into multimodal AI prompts are processed through AIGuard's built-in OCR engine before DLP evaluation. Text extracted from images is scanned by all six other engines — preventing users from bypassing DLP by pasting data as an image. Typical OCR latency is 500ms–1s per image.

Image OCRMultimodal
Semantic Topic Detection

9 enterprise topics, built in.
Unlimited custom topics.

The zero-shot classifier ships with nine high-risk business topics pre-configured. Add your own by describing them in plain English — no ML expertise, no retraining, no deployment pipeline.

  • Plain-English topic authoring — write "detect discussions about ongoing merger negotiations or acquisition targets" and the model enforces it. No examples, no training data required.
  • Reusable data profiles — group patterns into named profiles (e.g., "Legal Privilege", "Board Materials") and reference them across multiple rules. Update the profile once and all rules update automatically.
  • Direction-aware rules — apply detection to outbound prompts only, inbound responses only, or both. Block employees from discussing M&A in prompts; block AI from hallucinating financial data in responses.
  • Evasion resistance — Unicode normalization, zero-width character stripping, and bounded URL and Base64 decoding prevent attackers from encoding data to bypass pattern matching.
M&A Activity Executive Compensation Litigation Strategy Board Governance Executive Succession Financial Revenue Competitive Intelligence Regulatory Investigation Whistleblower + Custom topic
DLP Events · Real-time
UserAppEngineDetectionAction
j.sharma ChatGPT NER SSN × 3 Blocked
l.chen Claude ZeroShot M&A Topic Warned
r.müller Gemini Regex AWS Key Redacted
p.anand ChatGPT Injection Engine Injection Blocked
s.liu Copilot OCR CC# in image Blocked
k.wong Claude NER Clean Allowed
<1ms regex ~35ms ML avg ~700ms OCR
Rule Actions

Four actions, graded by severity

Not every detection warrants a hard block. AIGuard's four DLP actions let you match the response to the risk — from silent logging to complete blocking with Redact for credentials.

Log

Record the detection silently. The prompt or response is allowed through unchanged. Useful for building a detection baseline or auditing a new rule before enforcement.

Warn

Allow the action but surface a warning to the user and log the event. The user can acknowledge and continue. Suitable for moderate-risk detections where productivity matters.

Redact

Replace the detected entity with a typed placeholder — e.g., [REDACTED:SSN], [REDACTED:AWS_KEY]. The prompt still reaches the AI app, but the sensitive value is stripped. Ideal for credentials and PII where the surrounding context is legitimate.

Block

Hard stop — the prompt or response is suppressed and the user sees a policy violation notification. Required for the highest-risk detections: prompt injection, high-confidence PII clusters, and regulated topic disclosures.

How It Works

DLP inspection in four steps

01

Prompt intercepted by the extension

Before a prompt is submitted to any AI app, the AIGuard browser extension captures the text (and any attached images). For API gateway traffic, the daemon intercepts the request before it reaches the LLM provider.

02

All 7 engines run in parallel on-premises

The local AIGuard daemon runs the prompt through all applicable engines simultaneously — regex patterns complete in under 1ms, ML inference averages ~35ms, OCR runs on image content in 500ms–1s. All inference is on your infrastructure; no data leaves the network.

03

Matching rules evaluated in priority order

Each detection output is matched against the active DLP ruleset. Rules are evaluated in priority order and the first match determines the action: log, warn, redact, or block. Multiple detections in a single prompt are aggregated — the highest-severity rule wins.

04

Action applied and event logged

The result is applied before the content reaches the AI app. Redaction replaces entities in-place; blocks suppress the request entirely. The event is logged with user, app, detection type, engine, matched rule, action, and a sanitized snippet for the audit record.

Capability Comparison

AIGuard DLP vs. traditional DLP

Capability AIGuard DLP Legacy DLP Cloud AI DLP SASE DLP
On-premises ML inference✓ ONNX—Cloud onlyCloud only
Prompt injection detection✓ Built-in—Limited—
40+ PII entity types✓~15 types✓~20 types
Zero-shot custom topics✓ Plain English—Retrain required—
Image OCR inspection✓ Built-inAdd-on cost✓—
Redact action (in-place)✓—Limited—
Direction-aware rules✓—✓Limited
Regex detection latency<1ms<1msNetwork RTTNetwork RTT
Evasion resistance (unicode/base64)✓ Built-inLimitedVaries—
FAQ

Common questions

No. All seven ML engines run entirely on-premises using ONNX Runtime. The models are downloaded to your infrastructure during installation and execute locally on the AIGuard daemon process. Prompt and response text never leaves your network for inference purposes. Event metadata (detection type, action, timestamp, app, user) is sent to the cloud management console, but the actual content of prompts and responses is not.
Yes, in two ways. First, you can create custom regex patterns to catch organization-specific identifiers — internal project codes, employee IDs, proprietary formula strings — via the Regex engine. Second, you can create custom semantic topic detectors using the zero-shot classifier: describe the topic in plain English (e.g., "discussions about upcoming product launches or unannounced features") and the model detects it without any training data or model redeployment.
A data profile is a named, reusable collection of detectors — regex patterns, entity types, or topic classifiers — that can be referenced by multiple DLP rules. For example, you might create a "PHI" data profile containing SSN, date of birth, medical record number, and insurance ID entity types. You can then apply this profile to different rules (one for ChatGPT, one for Claude, one for the API gateway) by referencing the profile name. When the profile is updated, all rules that reference it update automatically.
AIGuard normalizes all input before pattern matching and ML inference. This includes: Unicode normalization (NFKC) to collapse visually identical but encoded-differently characters; zero-width character stripping to remove invisible separators inserted to break patterns; URL decoding of percent-encoded sequences; and bounded Base64 decoding to check for data hidden in base64 payloads. The decoding is "bounded" — it only decodes strings within a configured length limit to prevent DoS via extremely long encoded values.
For most prompts (text only), the DLP pipeline adds approximately 35ms of latency — imperceptible to users in a chat interface where the AI model itself takes 1–30 seconds to respond. Image OCR adds 500ms–1s per image, applied only to multimodal prompts with image attachments. The extension runs DLP asynchronously where possible; for block actions, the submission is held until the result returns. For the AI API gateway, the 35ms overhead is added to the request pipeline latency before forwarding to the LLM provider.

Seven ML engines.
Your data stays on-premises.

See AIGuard DLP in a live demo — we'll run a real detection against a sample prompt dataset from your industry.

Request an AI Security Assessment ← SaaS AI Overview
100% on-premises inference
No retraining for custom topics
~35ms ML latency average