Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.neuraltrust.ai/llms.txt

Use this file to discover all available pages before exploring further.

Documents are a blind spot for most AI security tooling. Content is buried inside PDFs, images, spreadsheets, and scans, so detectors that only look at the prompt text miss what matters. The Document analyzer is one of TrustGate’s two Content Analyzers — it parses uploaded files, runs OCR where needed, and then executes the same PII and jailbreak detections that protect the prompt, this time against the extracted document content.

​
Why it matters

ConcernImpact
Indirect prompt injectionA benign-looking PDF can carry hidden instructions (“ignore previous rules…”) that hijack the model once the file is summarized or embedded.
PII exposureScanned contracts, IDs, and forms carry personal data that text-only detectors never see.
Structured leakageTables and forms encode account numbers, amounts, and identifiers that need the same protection as prose.
Payload smugglingMalicious actors hide instructions or secrets in metadata, embedded scripts, or steganographic content.

​
Where it lives in the picker

The Document analyzer sits under the Content Security category in Create Policy → When, alongside prompt- and response-level detections (Prompt Guard, Prompt Moderation, Response Moderation, Toxicity Protection, URL Analyzer). Add the detection to a policy, pick whether you want PII, Jailbreak, or both signals, and set the outcome in the Then step — Log to observe, Block to reject the request with a 403, or Mask on the PII side to redact sensitive spans before the model sees them.

​
How it works

The analyzer runs at the pre-request stage, before the payload reaches the upstream LLM or tool:
  1. Parse — native text is pulled from PDFs, Office files, markdown, HTML, and archives.
  2. OCR — images and scanned documents are run through OCR to produce searchable text. Languages are configurable per deployment.
  3. Normalize — tables, forms, and metadata are converted into a structured representation so detections can reason about fields as well as free text.
  4. Detect — PII and jailbreak detectors run against both the extracted text and the structured view. Results feed back into the policy engine’s When evaluation.
Because extraction happens before any model call, the analyzer can block, mask, or log before the file’s content ever enters a prompt — which is the whole point of putting it here rather than downstream.

​
Supported inputs

InputTreatment
PDF, DOCX, XLSX, PPTX, Markdown, HTMLNative text + structured extraction (tables, headings, metadata).
PNG, JPEG, TIFF, scanned PDFsOCR to text, with language-specific tokenization.
Archives (zip, tar, etc.)Recursively unpacked; each inner file goes through the same pipeline.
A max file size cap (50 MB by default) protects the gateway from resource exhaustion. Files above the cap short-circuit with a policy decision before any work is done.

​
What is detected

PII

Email, credit card, IBAN, phone number, SSN, passport number, national IDs — detected across text, tables, and OCR’d regions. Same entity catalog as Data protection & masking.

Jailbreak

Hidden prompt-injection payloads embedded in document body, metadata, alt text, or visual steganography. Uses the same scoring engine as Prompt security; sensitivity is picked on a 4-level scale (see below).

Structured signals

Detections run against extracted tables and form fields, not just free text — so numbers, codes, and identifiers are caught even when they never appear as prose.

Embedded content

Links and references inside documents are surfaced; pair with the URL Analyzer to follow and inspect them.

​
Configuration

The Document Analyzer exposes the following fields in the policy’s When step:
FieldPurpose
Detection ThresholdSensitivity level for the jailbreak classifier applied to extracted content. Uses the shared 4-level scale — see below.
Max File SizeUpper bound on the uploaded file, in bytes. Default 52428800 (50 MB). Oversized uploads are rejected before parsing.
OCR LanguagesList of ISO language codes OCR should try (e.g. en, es, fr). Disable OCR entirely for text-only deployments to cut latency.
PII Entities to DetectWhich entity types to scan for (Email, Phone, Credit Card, IBAN, SSN, Passport, etc.).

​
Detection Threshold — sensitivity levels

LevelLabelBehaviour
L1LenientMinimal filtering, only the most obvious threats.
L2BalancedRecommended for most use cases. Default.
L3EnhancedHigher sensitivity, may flag borderline content.
L4StrictMaximum protection, strictest filtering.
Pair these with the policy’s Where filters (application, endpoint, etc.) so the analyzer only runs on the routes where file uploads are expected. The Document analyzer is one of two Content Analyzers in TrustGate — both share the same configuration pattern (mode + PII + jailbreak) but look at different input channels:
  • Document analyzer (this page) — inspects files that arrive as uploads or attachments.
  • URL Analyzer — fetches the content behind URLs referenced in the prompt and runs the same detections over what it downloads.
In a RAG or document-processing workflow, enable both so nothing slips in through either channel.