angusdevgo / Seep-Reverse-Lab
git clone https://github.com/angusdevgo/Seep-Reverse-Lab.git
Agent-Native multi-platform reverse engineering and CWE-602 client-side authorization audit workbench.
β
stars, 24h
β
stars, 7 days
+1.7k
stars, 30 days Β· #179
1.7k
β
total
636
forks Β· 1 open issues
Star history starts 2026-10-20; the chart appears after the second daily snapshot.
LanguageLua
LicenseGPL-3.0
Created2026-09-23 (27d ago)
Last push1d ago
Watchers3
OwnerUser
Links
README excerpt
Seep Reverse Lab Agent-Native Β· Multi-Platform Reverse Engineering Β· CWE-602 Authorization Audit Β· Autonomous Security Workbench [ English | δΈζζζ‘£ ] Overview β’ Architecture β’ Demo β’ MCP Matrix β’ Quick Start β’ Scope β’ Agent Contract β’ Disclaimer --- > π **Attribution & Reference Sources**: > - Mobile reverse engineering methodology & verification test suite referenced from: **newliver666/apk-reverse** (MIT License). > - Security research lab framework, Zero-Waste Recon signal routing & agent execution protocols referenced from: **GeniusHu-tgty/Open-tgtylab** (GPL-3.0 License). > - Reverse engineering knowledge base, attack graph boards & MCP automation ecosystem referenced from: **LING71671/open-reverselab** (GPL-3.0 License). > - Community support & technical discussions: **LINUX DO**. --- ## π Project Overview Seep consolidates fragmented reverse engineering toolchains (Radare2 / JADX / Apktool / Frida / IDA), operational knowledge bases, prompt engineering contracts, and battle-tested field experience into an **Agent-Native Autonomous Workbench**. Users provide plain-language technical goals; the agent autonomously performs platform identification, vulnerability type classification, surgical binary patching, and closed-loop evidence delivery. ### Five Core Pain Points Solved | Pain Point | Solution | |---|---| | π§ **Toolchain Fragmentation** | 23 MCP tools wrapping Radare2, JADX, Apktool, Frida β unified API, zero manual switching | | π§ **Agent Decision Drift** | `softseep` orchestrator: two-stage auto-classification (Platform Γ 9 Task Types) + 7-gate decision tree | | π‘οΈ **Model Refusal in Security Audits** | Three-tier mitigation: `BLOCK_NONE` injection + transparent terminology mapping + cognitive redirection | | π **Ephemeral Knowledge** | 289 technical field journals + 14 desensitized paradigms + "search KB before executing" enforced discipline | | π¦ **Setup Friction** | One-click `install.ps1` β supports Pi Agent, Claude Code, DeepSeek Harness (DSH), Codex / OpenCode | --- ## βοΈ How It Works [code] --- ## β‘ 30-Second Demo Once deployed, just talk to your agent in plain language: [code] The agent autonomously: 1. **`seep_auto_triage`** β identifies DEX + ARM64 SO, no packer detected 2. **`seep_apk_decompile`** with `output_mode=fold` β extracts control-flow skeleton only *(saves ~60% tokens)* 3. **`seep_kb_search`** β finds matching CWE-602 pattern from field journal #142 4. **`seep_apk_gen_hook`** β generates `hook_verify.js` Frida script 5. Executes on connected device β captures stdout, **self-heals** `ClassNotFoundException`, re-runs 6. Airplane-mode confirmation β β **`seep_gen_security_report`** β 3-part audit report delivered > Total elapsed time on a typical client app: **8β20 minutes**, fully unattended. --- ## β‘ Core Capabilities - π― **Natural Language Intent Resolution**: Binary magic bytes + semantic verb mapping β correct toolchain, no flags to memorize - π **CWE-602 Authorization Audit**: Minutes to determine