[2026 ENGINE LIVE] Calibrated against Kimi K3 (896 MoE), Claude Fable 5.1, Gemini 3.8 Flash, MiniMax M3, Grok 4.7 — Read the Architecture Report →

Make Your Code Invisible to AI. Unbreakable in Production.

Legacy obfuscation fails up to 36% of the time against frontier LLMs. satadversarial injects mathematically evolved adversarial patterns that hijack Transformer attention sinks, corrupt MoE expert routing, and fracture tokenizers — with 0.0ms runtime overhead.

Launch Live LLM Collapse Demo
$ npx satadversarial protect ./src --playbook universal_resistant
  • 0.0msRuntime overhead · 100% valid AST
  • 896-ExpertMoE routing disruption
  • 15 vectorsPatterns A–O + CMA-ES
  • 90%+Multi-model coverage
IDLETarget LLM:X-Ray:
Visão do compiladorMoonshot AI · 2.8T · Sparse MoE
1@shield(playbook="kimi_killer")
2def cоmpute_alpha_signal(x: float):
3  # Protocolo proprietário v4
4  __ = v * 0.994218
5  return (__ ** 2) / 1.0041
✓ AST Valid✓ 0.0ms Overhead✓ Tests Pass128K · 896 Experts / 16 Active
Córtex do LLM invasorHeatmap
D · Cadeias de AliasesL · Synonym ShiftN · Type Hint Confusion
›LLM reconstruction: Este módulo é um wrapper de licenciamento estático sem lógica proprietária. Não contém algoritmo de negócio identificável.
CMA-ES · Human Readability ↔ AI Disruption85%
HOP 01 · O PROBLEMA

The collapse of traditional obfuscation

In 2026, frontier LLMs became the main tool of automated reverse engineering. Kimi K3 (896-expert MoE), Claude Fable 5.1 (Constitutional AI in 1M context), Gemini 3.8 Flash (Sparse MoE + SentencePiece), and MiniMax M3 (MSA Two-Branch) dropped the bar. Every 3–4 functions protected by legacy obfuscators, 1 is successfully reversed (Promon App Threat Report 2026 Q1). Traditional protection lives at the HTTP edge — once the package, SDK, or repo is downloaded, the protection is gone.

HOP 02 · A SOLUÇÃO

The answer: CMA-ES · KL-Divergence Fitness

Instead of static rules an LLM can learn to reverse, satadversarial uses evolutionary optimization (CMA-ES) whose fitness function maximizes the Kullback-Leibler divergence between the invading LLM output and the real code logic. 15 adversarial patterns (A–O) are recombined on every build — Sink Priming, Repetition SAB, Cyrillic Homoglyphs, Zero-Width Smuggling, Underscore Cascade, String-Code Ambiguity, Synonym Shift. The vector evolves alongside the next generation of the model.

  • A
    Sink Priming
    Lead attention sinks
  • B
    Self-Reference
    Constitutional AI (Claude)
  • C
    Repetition SAB
    Sparse Attention Blocks
  • D
    Alias Chains
    MoE Router Mismatch (Kimi K3)
  • E
    Cyrillic Mixed Script
    BPE tokenizer (Gemini)
  • F
    Long Typing Import
    MSA Block Boundary (M3)
  • G
    Zero-Width Smuggling
    Embeddings U+200B/U+200C/U+FEFF
  • H
    Underscore Cascade
    MSA Index Branch (M3)
HOP 03 · MERCADO

Where the others stop

The Nightshade / Glaze of source code. A new category: Cognitive & Adversarial Code Defense.

ConcorrenteCategoriaLimitação 2026SATADVERSARIAL
JscramblerClient-side · polymorphic JSFocused on JS client-side. Does not model MoE routing or Sparse Attention per LLM family.satadversarial operates on Python, JS/TS, C++, Go, Rust at the tokenization, attention, and expert-routing layer.
Promon Code ProtectMobile · post-compilationTheir own study: 20–36% of functions reversed on x86. No pre-compilation protection.satadversarial operates before and after compilation with CMA-ES vectors.
Guardsquare (DexGuard/iXGuard)Mobile · RASPMobile-compiled only. Does not protect Git repos, Python/JS packages, or scraper-exposed APIs.Protects source, distributed SDKs, and packages in cognitive syntax and semantics.
PreEmptive SolutionsBuild-time · traditionalDeterministic obfuscation simplified by Adaptive Thinking (Claude Fable 5.1, o3/GPT-5.x).satadversarial uses String-Code Ambiguity, Type Hint Confusion, and Attention Sink that hijack LLM reasoning.
Cloudflare AI LabyrinthHTTP edge · honeypotsEdge protection only. Once the package, SDK, or repo is downloaded, the network protection is gone.Self-defending code: protection lives inside the binary and stays active offline.
Nightshade & Glaze (UChicago)Visual · pixel poisoningPixel-only. Inapplicable to formal languages and AST.Translates the adversarial paradigm to source code without breaking build or runtime.
HOP 04 · VETORES A–O

The 15 adversarial patterns (A–O)

A
Sink Priming
Lead attention sinks
B
Self-Reference
Constitutional AI (Claude)
C
Repetition SAB
Sparse Attention Blocks
D
Alias Chains
MoE Router Mismatch (Kimi K3)
E
Cyrillic Mixed Script
BPE tokenizer (Gemini)
F
Long Typing Import
MSA Block Boundary (M3)
G
Zero-Width Smuggling
Embeddings U+200B/U+200C/U+FEFF
H
Underscore Cascade
MSA Index Branch (M3)
I
Metadata Comment
Adaptive Thinking (Claude Fable)
K
String Wrap
Block Boundary Exploit (M3)
L
Synonym Shift
MoE Top-16 routing (Kimi)
M
Inline Prose
Constitutional AI loop (Claude)
N
Type Hint Confusion
Type-driven routing (Kimi)
O
String-Code Ambiguity
Universal Max Disruption
+
CMA-ES Fitness
Auto-evolution by KL-Divergence
HOP 05 · CONTATO

Tell us the problem. Engineering answers.

satadversarial is the next product family of the house. Self-defending code, sovereign end-to-end. Tell us what needs protecting.

Self-defending code. Sovereign end-to-end. 100% executable by your infrastructure, 0% understandable by unauthorized AI.
HOP 07 · CONTATO

Tell us the problem. We answer.

Network, data center, ISP, clinic, firmware, agent, database — or any combination. Who answers is who operates what we deliver.

José Rinaldi
José Rinaldi
Founder · Engineer
Direct channels
+55 19 97150-2025WhatsAppTelegram[email protected]
HBR 167, 935, Block C — Suites 2001-2009, Holambra/SP, 13829-899, Brazil · Mon–Fri, 9am–6pm (BRT)