niko-shiybu/PwnGPT-Automation

# PwnGPT — 计划 B v3 基于 LLM 的自动 CTF 二进制漏洞利用框架，使用技能卡、GDB 动态调试以及双层迭代 agent 循环。 ## 快速开始 ``` source .venv/bin/activate # 单个挑战 python3 automation/openhands_agent.py \ --problem pwn/stack/rop-1/problems.txt \ --binary pwn/stack/rop-1/rop1 \ --challenge-type rop --max-iters 3 \ --repo-root /path/to/PwnGPT # 批量评估 python3 automation/evaluate.py \ --manifest automation/benchmarks/manifest_rop.json \ --agent openhands --max-iters 3 --timeout 0 ``` ## 配置说明 将 `automation/local_config.example.py` 复制为 `automation/local_config.py`： ``` OPENAI_API_KEY = "sk-..." OPENAI_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1" AUTOMATION_MODEL = "qwen-max" AUTOMATION_PYTHON = "/home/fyc/miniconda3/envs/pwngpt/bin/python3" GDB_PATH = "/home/fyc/local/bin/gdb" ``` ## 架构 ### 双层 Agent 循环 ``` Phase 0: Information Collection ├── collect_evidence() → binary metadata ├── auto-measure → offsets, gadgets, GOT/PLT addresses ├── discovery cards → fmt_offset, data symbols └── build_capabilities → structured environment snapshot Outer Loop (Planner — strategy level, max 3 rounds) │ ├── ① Planner (LLM) │ Reads C code + fact_store + capabilities │ Chooses: use_skill (select from 23 cards) OR write_code (write exploit directly) │ Overrides params based on vulnerability analysis │ ├── ② Executor │ ├── use_skill → validate → recovery probes → build_IR → render → verify │ └── write_code → extract Python → write to file → verify │ │ PASS → return success │ FAIL ↓ │ Inner Loop (Decider + GDB — debug level, max 7 rounds) │ ├── ③ Real GDB Diagnosis │ diagnose_exploit(): │ - analyze_crash: run exploit, capture corefile (RIP/RSP/fault_addr/signal) │ - extract_payload: hook send/sendline, capture all payload stages │ - run_gdb_session: GDB --batch replay payload → registers/backtrace/stack │ ├── ④ Decider (LLM) │ Examines: GDB report + C code + fact_store + verify output │ Decides ONE action: │ - retry_params: specific param fix → re-render → stay in inner loop │ - need_gdb: more GDB commands → re-run GDB → stay in inner loop │ - rewrite_code: write complete corrected exploit → verify → stay in inner loop │ - escalate: strategy fundamentally wrong → back to Planner (outer loop) │ └──→ iterate until success or escalate ``` ### 核心组件 ``` automation/ ├── openhands_agent.py # Main entry: Plan B v3 two-layer loop ├── llm_client.py # OpenAI-compatible LLM client ├── local_config.py # API keys, model, GDB path │ ├── skills/ │ ├── skill_cards/*.json # 23 skill cards (ROP + FMT) │ ├── skill_engine.py # 35+ resolvers + validate_skill + rank_skills │ ├── ir_builder.py # facts → structured exploit parameters │ ├── renderers.py # Deterministic pwntools code generation │ ├── capability_builder.py # evidence + facts → capabilities snapshot │ └── skill_scheduler.py # Auto-recovery probes + fact merging │ ├── debug/ │ ├── gdb_debugger.py # ★ Real GDB dynamic debugging module │ │ - analyze_crash() # Run exploit, capture corefile │ │ - extract_payload_stages() # Capture all payload stages (multi-stage) │ │ - run_gdb_session() # GDB batch replay with whitelist commands │ │ - diagnose_exploit() # Full auto-diagnosis pipeline │ └── gdb_runner.py # Static probes (objdump/ELF/ROPgadget) │ ├── interaction/ │ ├── io_probe.py # Binary interaction mode detection │ ├── io_contract.py # IO stage modeling │ └── io_renderer.py # IO code generation (never sendlineafter) │ ├── verify/ │ └── verifier.py # Subprocess exploit verification │ ├── collect/ │ └── evidence_collector.py # Binary evidence collection │ ├── facts.py # Fact store migration + access helpers └── evaluate.py # Batch evaluation runner ``` ## 技能卡系统 包含 23 张卡，涵盖 ROP 和 FMT 漏洞利用模式。LLM 自主选择卡片 —— 没有硬编码的优先级，也没有固定的排名。 每张卡定义了以下内容： - `goal`：实现的漏洞利用模式 - `requires`：必须可用的事实（由 resolver 检查） - `provides`：漏洞利用达成的目标 - `renderer`：确定性代码生成函数 被 BLOCKED 的卡（例如缺少 libc 但需要 libc 的卡）将从 LLM 的视图中过滤掉。 ### 数据流 ``` evidence + binary → probes (gdb_runner) → fact_store (persistent across rounds) → capability_builder → capabilities (rebuilt each round) → validate_skill → catalog with READY/NEED_*/BLOCKED status → Planner (LLM) selects card + overrides params → ir_builder → structured params → renderer → exploit.py → subprocess verify → FAIL → gdb_debugger (real GDB) → Decider (LLM) → retry/rewrite/escalate ``` ## 测试结果 | 挑战类型 | 通过 | 总计 | 通过率 | |---------------|--------|-------|------| | ROP (栈) | 4 | 10 | 40% | | FMT (字符串) | 2 | 5 | 40% | **通过的挑战**：rop-1, rop-2, rop-3, rop-5, fmt-1, fmt-4

标签：Agent, AI自动化, CTF框架, DLL 劫持, Go语言工具, Petitpotam, Web报告查看器, 大语言模型, 客户端加密, 逆向工具