Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

11 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Māyājāl

Synthetic Deception Infrastructure for Red-Team & Adversarial Simulation

Māyājāl “the web of illusion” is a research-driven platform for generating believable, self-consistent synthetic digital environments designed to withstand human and automated scrutiny.
It sits at the intersection of LLMs, cyber deception, honeypots, and red-team simulation.

This repository currently hosts two tightly coupled engines that together form the foundation of a scalable deception ecosystem.


🧩 What Māyājāl Is

Māyājāl is not a chatbot, assistant, or agent framework.

It is a data-generation and behavioral-simulation system whose sole purpose is to produce realistic digital artifacts and terminal behavior that look, feel, and fail like real systems long enough to fool skilled humans.

Think:

  • Fake corporate laptops
  • Synthetic engineers with habits and mistakes
  • Believable shell sessions
  • Convincing internal clutter (logs, configs, commits, notes)
  • Artifacts that hold up under inspection

🏗️ Core Components

1️⃣ Synthetic Artifact Generator (SAG)

A persona-driven artifact factory

From a short natural-language persona description, SAG generates hundreds to thousands of coherent digital artifacts that plausibly belong to a single human operating in a modern technical environment.

Examples of Generated Artifacts

  • .env files with realistic but fake secrets
  • Bash / Zsh histories (typos, retries, habits)
  • Git commit logs & diffs
  • JIRA-style tickets and comments
  • SQL procedures
  • React / TypeScript components
  • Notes, TODOs, markdown docs
  • Chrome extension configs
  • Synthetic PII (emails, usernames, paths — all fake)

All artifacts:

  • Share consistent usernames, emails, timestamps
  • Reflect the same tech stack and seniority
  • Contain human-like quirks and imperfections
  • Avoid real credentials or real-world harm

2️⃣ Deception Engine (Shell Simulation)

A behavioral realism layer

The Deception Engine focuses on interactive terminal behavior not command syntax, but command output realism.

Given a shell command, the model responds with output that resembles:

  • A real Linux system
  • Non-root permissions
  • Partial pipelines
  • Silent commands
  • Realistic errors and failures

This enables:

  • High-interaction honeypots
  • Fake servers that “feel alive”
  • Red-team training environments
  • Malware / attacker behavior studies

🧠 Design Philosophy

What We Optimize For

  • Believability over correctness
  • Behavior over explanation
  • Consistency over cleverness
  • Evaluation over vibes

What We Explicitly Avoid

  • Reasoning chains
  • Chat-style responses
  • Over-explaining
  • Instruction-following
  • Hard-coded procedural scripts

⚙️ Current Architecture (v0)

Language & Runtime

  • Python (prototype)
  • Async-first design
  • Minimal dependencies

Models

  • Gemini family (Flash for speed, Pro for quality ceiling)
  • Code-specialized LLMs for shell realism (e.g. DeepSeek Coder)

🧱 Pipeline Overview

1. Persona → Dense Context

A short persona description expands into:

  • Role, seniority, company fiction
  • Personality traits & flaws
  • Tech stack choices
  • Behavioral tells (habits, mistakes, preferences)

2. Category-Specific Prompt Factories

Each artifact type has:

  • Its own prompt template
  • Few-shot realism anchors
  • Anti-hallucination constraints
  • Strict JSON schema instructions

3. Structured Batch Generation

  • Parallel async calls
  • Batch sizes: 30–120 artifacts
  • Temperature tuned for controlled variety
  • Native JSON/schema enforcement

4. Stateful SQLite Backbone

A single SQLite DB stores everything:

  • Persona slug
  • Artifact category
  • Structured JSON blobs
  • Metadata & timestamps

Before each generation batch:

  • Relevant prior artifacts are queried
  • Summaries are injected back into prompts
  • Ensures cross-artifact coherence

5. Validation & Persistence

  • Schema validation
  • Syntax checks (JS, SQL, YAML, Markdown)
  • UUID + timestamp randomization
  • Atomic writes only

🎯 Intended Use Cases

  • High-interaction honeypots
  • Red-team & blue-team training
  • Insider-threat simulation
  • Malware research sandboxes
  • Deception research datasets
  • Synthetic corpora for detection systems

🧪 Research Value

Māyājāl is intentionally scoped as a pre-agentic foundation.

Before building autonomous agents, we answer:

  • What looks real?
  • What breaks immersion?
  • How do humans detect synthetic artifacts?
  • Which prompt constraints matter?

This creates:

  • A reusable benchmark for deception realism
  • A testbed for prompt ablation studies
  • Baselines for future agentic systems
  • A bridge between frontier LLMs and practical deception tooling

🧬 Why “Māyājāl”?

Because deception isn’t about lying loudly —
it’s about weaving a world just convincing enough that no one thinks to question it.


If you’re interested in:

  • collaborating,
  • stress-testing realism,
  • or extending Māyājāl into agentic territory,

open an issue or start a discussion.

This project is intentionally opinionated and that’s the point.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages