AgenticART¶

Verifiable Security Research & Dynamic Analysis

The Problem¶

LLMs generate exploit code that looks correct but doesn't run:

• Uses APIs that don't exist (frida.hooks.Hook)
• Invents kernel structures and syscalls
• Never receives execution feedback

Root cause: Models pattern-match syntax without knowing what actually executes.

The Solution¶

AgenticART creates a Praxis Loop: a reasoning-to-verification feedback loop between the model and a real Android device:

flowchart TD
    A[Artifacts: APK/Manifest] --> B[Reasoning: OBSERVE/HYPOTHESIZE]
    B --> C[Verification: Execute MCP Tools]
    C --> D{Calibration OK?}
    D -->|Hallucination| E[Self-Correction]
    E --> B
    D -->|Verified| G[DPO Training Data]

Failures become intelligence. The model learns to match its confidence to reality through empirical verification.

Key Results¶

Belt Progression¶

Models advance through structured difficulty levels:

Requirements¶

Documentation¶

Research¶

Inspired by "LLM-Powered Android Exploitation" which introduces the feedback loop methodology.