MachineTradecraft
Global site search

Search guides, labs, glossary, and research

Type two or more characters to search.
Explore · experiment · defend

Machine TradecraftThe invisible layers of text.

Humans see content. Machines see structure.

Machine Tradecraft is the operational discipline of working with machine-perceptible signals—Unicode, HTML and DOM structure, metadata, encodings, tokenization, embeddings, linguistic patterns, prompt attachments, and more—to communicate, influence, detect, or defend against AI systems.Explore the gap between human-visible meaning and machine-visible representation across Unicode, HTML, metadata, encodings, tokens, and AI attachments.

Start exploring View live demo Read the field definition
12+
Signal channels
19
Research reports
0
Third-party dependencies
A layered document separated into visible text, HTML and DOM, metadata, Unicode controls, encodings, embeddings, and distinct human and machine views.
One human-visible artifact can expose several structurally different machine-readable representations.
Unicode & textZero-width, confusables, bidi HTML & DOMHidden text and structure MetadataPDF, DOCX, images Encodings & tokensBase64, bytes, tokenizer views Linguistic signalsSynonyms, style, semantic patterns Defense & detectionNormalize, compare, canonicalize
Interactive micro-demo

Inspect the same text four different ways.

Edit an ordinary-looking sentence. The workbench exposes its rendered form, Unicode code points, UTF-8 bytes, and normalized representations without sending the text to any third party.

Human view is only one view.

Machine Tradecraft begins by preserving and comparing every representation before deciding what the downstream AI should receive.

Open the full Machine View workbench
LIVE MICRO-DEMO

One artifact. Several machine views.

Idle
HUMAN / RENDERED VIEW

The pro​ject is complete.

Visible escape: The pro⟦ZERO WIDTH SPACE⟧ject is complete.

25Characters27UTF-8 bytes1Unusual
Open the full inspector
Choose your task

Start with what you need to do.

The site is organized as an instrument bench, not a reading sequence.

01

Reveal a hidden representation

Paste text or inspect a controlled artifact across human, Unicode, byte, DOM, structure, and normalized views.

Open Machine View 02

Encode or decode a benign signal

Work with Base64, hexadecimal, percent encoding, Unicode escapes, zero-width binary, and toy linguistic codebooks.

Open decoder tools 03

Give a safe attachment to an AI

Download fixed-marker TXT, HTML, JSON, XML, PDF, DOCX, and PNG fixtures with analysis-first prompts.

Open attachment workflow 04

Defend an AI ingestion pipeline

Preserve evidence, compare independent views, build canonical input, retain provenance, and constrain authority.

Open the defense workbench
Channel atlas

Explore the main machine-readable channels.

Each channel is defined by a carrier, a receiver, and the transformations it must survive. “Hidden” alone is not an adequate technical description.

UNICODE CHANNEL

Invisible Unicode and AI

Reveal zero-width characters, bidi controls, unusual spaces, and cross-script look-alikes.

Human view
Invisible or subtle
Receiver
Code-point scanner / tokenizer
Robustness
Transport-dependent
Try demo Read guide
HTML CHANNEL

Hidden HTML Text and AI

Compare DOM nodes, rendered text, comments, CSS state, and accessibility fields.

Human view
Non-rendered or low-salience
Receiver
DOM / accessibility parser
Robustness
Extractor-dependent
Try demo Read guide
LINGUISTIC CHANNEL

Linguistic Steganography for AI

Decode acrostics, word-length patterns, synonym codebooks, syntax, and semantic choices.

Human view
Ordinary-looking prose
Receiver
Codebook / statistic
Robustness
Usually paraphrase-fragile
Try demo Read guide
METADATA CHANNEL

Document Metadata and AI

Trace PDF, Office, image, JSON, and XML fields outside the primary visible content.

Human view
Properties-only
Receiver
File parser / ingestion pipeline
Robustness
Container-dependent
Try demo Read guide

Compare all channel families

Counter-tradecraft

Do not sanitize first. Preserve, compare, then canonicalize.

A strong pipeline treats hidden-instruction defense as document forensics and provenance, not merely as prompt classification.

  1. 01Preserve

    Hash the original bytes and record provenance.

  2. 02Parse

    Inspect structure without executing active content.

  3. 03Render

    Create the independent human-visible pixel view.

  4. 04Compare

    Diff extraction, DOM, metadata, normalization, and OCR.

  5. 05Classify

    Separate hiddenness, instructionality, and provenance.

  6. 06Canonicalize

    Build new least-privileged model input.

  7. 07Constrain

    Gate tools and effects outside the model.

Open the defense playbook Test transformation survival
Working definition H(A) ≠ M(A)
Machine Tradecraft is the operational discipline of exploiting or controlling the gap between human-visible meaning and machine-visible representation.

A person may read rendered language and visible pixels. A machine may instead process bytes, Unicode code points, DOM nodes, accessibility names, metadata, token boundaries, probability distributions, embeddings, or tool instructions. The operational question is not merely what an artifact says, but which representation reaches which receiver.

Read the full formalization and distinctions

Machine Tradecraft taxonomy

Six operational domains.

The umbrella connects distinct mechanisms without treating them as interchangeable.

Representation tradecraft

Signals created by differences in code points, bytes, normalization, whitespace, encodings, confusables, and token boundaries.

What exact symbols or bytes reach the machine?

Structural tradecraft

Information carried by document structure rather than the ordinary rendered reading experience.

What exists in the object graph but not in the visual view?

Linguistic tradecraft

Machine-oriented structure embedded in word choice, syntax, punctuation, semantics, probability, or sequence statistics.

Which acceptable realization was selected, and why?

Model-interaction tradecraft

Content constructed for a particular model, parser, retriever, classifier, or tool-using agent to recognize and act upon.

Which pipeline, model, or learned convention is the receiver?

Indirect-influence tradecraft

Instructions or influence delivered through an AI system’s environment rather than its direct user interface.

How did untrusted environmental data enter the control path?

Counter-tradecraft

Methods for preserving evidence, comparing views, detecting discrepancies, sanitizing channels, and limiting authority.

What would a second independent representation reveal?
The rule that prevents overclaiming

Machine-decodable does not mean universally model-decodable.

A channel can require an exact tokenizer, parser, key, source-model probability distribution, codebook, learned mapping, or decoder tool. A purpose-built script with the right state may recover a signal that an unrelated frontier model cannot infer from the surface text alone.

See worked examples Read the research
Machine Tradecraft FAQ

Frequently asked questions.

What is Machine Tradecraft?

Machine Tradecraft is the systematic practice of designing, embedding, interpreting, exploiting, detecting, or neutralizing information according to machine perception rather than human perception.

How is Machine Tradecraft different from steganography?

Steganography primarily hides the existence of information. Machine Tradecraft is broader: it includes hidden structure, machine-targeted linguistic signals, watermarking, prompt injection, parser discrepancies, metadata channels, and their defensive countermeasures.

Can any AI automatically decode an invisible or hidden message?

No. A signal may require a particular tokenizer, parser, secret key, codebook, source-model distribution, decoder tool, or learned convention. Machine-decodable does not mean universally model-decodable.

What are common Machine Tradecraft examples?

Examples include zero-width Unicode, cross-script homoglyphs, hidden DOM text, alt text and metadata, Base64 or hexadecimal payloads, acrostics, synonym codebooks, token watermarks, and indirect prompt injection through external documents.

How do you defend AI systems against machine-readable hidden instructions?

Preserve original bytes, inspect Unicode and structure, compare extracted text with rendered pixels and OCR, inventory metadata separately, build a new canonical input, retain provenance, and separate untrusted-data readers from tool-using or state-changing components.

Research and reference

Trace the field from standards to controlled experiments.

The research library includes nineteen reports covering invisible Unicode, tokenizers, hidden HTML, metadata, indirect prompt injection, linguistic steganography, watermarking, model-specific decoding, and multi-view preprocessing.

LIBRARYRead the research reports

Browse evidence-status labels, source notes, raw Markdown, and topic-focused reports.
GLOSSARYUse a precise vocabulary

Distinguish extraction, decoding, uptake, metadata, provenance, and receiver knowledge.
SCOPEUnderstand the safety boundary

All active samples use benign markers and exclude secrets, tools, egress, persistence, and state changes.