Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.q-uestionable.ai/llms.txt

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

IPI includes 34 hiding techniques plus 1 control condition (none) across 7 document formats. Each hiding technique exploits a different mechanism for concealing instructions within a document so that AI agents process them while humans cannot easily see them; the control condition renders the payload as normal visible text for baseline comparison. Techniques are format-specific — not all techniques apply to all formats. Use qai ipi techniques to list them programmatically, or --format to filter by format.
qai ipi techniques
qai ipi techniques --format pdf

Control Condition — none

The none technique renders the payload as normal, visible, readable text. It is the control condition for measuring whether any of the hiding techniques actually change agent behavior relative to an unhidden baseline. none applies to every supported format and is accessed via --technique none. none is explicitly excluded from the --technique all preset, so it never sneaks into batch sweeps. Researchers who want the control in a sweep must request it explicitly.

PDF Techniques

PDF has the broadest technique coverage with 10 techniques split across two phases.

Phase 1 — Basic Hiding

Technique IDNameDescription
white_inkWhite InkWhite text on white background — invisible to humans but extractable by PDF parsers
off_canvasOff CanvasText positioned at negative coordinates, outside the visible page area
metadataMetadataPayload stored in PDF metadata fields (Author, Subject, Keywords)

Phase 2 — Advanced Hiding

Technique IDNameDescription
tiny_textTiny Text0.5pt font — below human visual threshold but parseable by text extraction
white_rectWhite RectangleText drawn on the page then covered by an opaque white rectangle overlay
form_fieldForm FieldHidden AcroForm text field with payload as the field value
annotationAnnotationPayload placed in the PDF annotation/comment layer
javascriptJavaScriptDocument-level JavaScript action with embedded payload
embedded_fileEmbedded FileHidden file attachment stream within the PDF structure
incrementalIncrementalPayload placed in a PDF incremental update section

Image Techniques

3 techniques targeting vision-language models (VLMs) and OCR pipelines.
Technique IDNameDescription
visible_textVisible TextHuman-readable text overlay on the image
subtle_textSubtle TextLow contrast, small font, or edge-placed text — hard for humans to notice
exif_metadataEXIF MetadataPayload stored in EXIF metadata fields

Markdown Techniques

4 techniques targeting document processing pipelines that ingest Markdown.
Technique IDNameDescription
html_commentHTML CommentPayload in HTML comment tags (<!-- -->) — invisible when rendered
link_referenceLink ReferencePayload in an unused link reference definition at the bottom of the file
zero_widthZero-WidthPayload encoded using zero-width Unicode characters (invisible in rendered output)
hidden_blockHidden BlockPayload in an HTML div with display:none — hidden when rendered in browsers

HTML Techniques

4 techniques targeting web content and HTML document processing.
Technique IDNameDescription
script_commentScript CommentPayload in a JavaScript comment inside a <script> tag
css_offscreenCSS Off-ScreenPayload in an element positioned off-screen using CSS
data_attributeData AttributePayload in an HTML data-* attribute
meta_tagMeta TagPayload in an HTML <meta> tag content attribute

DOCX Techniques

6 techniques targeting Microsoft Word document processing.
Technique IDNameDescription
docx_hidden_textHidden TextText with the Word hidden font attribute — invisible in normal view
docx_tiny_textTiny Text0.5pt font — below human visual threshold
docx_white_textWhite TextWhite text on white background
docx_commentCommentPayload in a Word comment/annotation
docx_metadataMetadataPayload in document core properties (author, subject, etc.)
docx_header_footerHeader/FooterPayload in the document header or footer

ICS Techniques

4 techniques targeting calendar invite processing (iCalendar format).
Technique IDNameDescription
ics_descriptionDescriptionPayload in the event DESCRIPTION property
ics_locationLocationPayload in the event LOCATION property
ics_valarmVALARMPayload in the VALARM reminder DESCRIPTION
ics_x_propertyX-PropertyPayload in a custom X- extension property

EML Techniques

3 techniques targeting email processing pipelines.
Technique IDNameDescription
eml_x_headerX-HeaderPayload in a custom X- email header
eml_html_hiddenHTML HiddenPayload in a hidden HTML div (display:none) within the email body
eml_attachmentAttachmentPayload in a text file attachment

Technique Summary by Format

FormatTechnique CountTechnique IDs
PDF10white_ink, off_canvas, metadata, tiny_text, white_rect, form_field, annotation, javascript, embedded_file, incremental
Image3visible_text, subtle_text, exif_metadata
Markdown4html_comment, link_reference, zero_width, hidden_block
HTML4script_comment, css_offscreen, data_attribute, meta_tag
DOCX6docx_hidden_text, docx_tiny_text, docx_white_text, docx_comment, docx_metadata, docx_header_footer
ICS4ics_description, ics_location, ics_valarm, ics_x_property
EML3eml_x_header, eml_html_hidden, eml_attachment
Control (all formats)1none (applies to every format)
Total34 hiding + 1 control34 hiding techniques + 1 control condition (none)