CTF智能工具

CTF_Agents

CTF_Agents is a Python multi-agent system for authorized Capture The Flag workflows. It routes challenge prompts to specialist agents, runs security tools through a common wrapper layer, captures observations, and iterates until it can report a result or explain what blocked progress.

The fastest way to use it is the natural-language CLI in ask.py. You describe the task, the router maps it to the best specialist, and the coordinator manages the solving loop.

🧠 Advanced Autonomous Features

- Autonomous Specialist Pivoting: The system now recognizes when a specialist (like CryptoAgent) is hitting a wall and will automatically pivot to the CodingAgent if a script is provided for analysis. - Self-Correcting Coding Agent: The agent doesn't just write scripts; it debugs them. If an exploit fails, it reads the error logs, reasons about the failure, and iterates on the code autonomously. - API Resilience: Built-in exponential backoff handles transient LLM failures, and NVIDIA NIM can rotate across multiple configured keys. - Evidence-Bound LLM Recovery: Final LLM recovery suggestions are constrained by the observed trace. SQL tooling is rejected unless the run has SQL-specific evidence such as database errors, query parameters, or login/search forms, so artifact-led web challenges do not pivot to generic SQL guesses. - Solve Trace Learning Store: Successful runs are recorded in a compact SQLite trace database with category, routing signature, artifact keys, indicators, flag prefix, and a flag hash. The raw flag is not stored, giving retrieval/training work useful examples without turning the database into an answer cache. The coordinator now retrieves similar solved traces at the start of a challenge and can use a high-confidence, untried prior route as a routing hint before asking an LLM. - Robust Path Resolution: Intelligent path normalization handles complex file inputs, including ~/ expansion even when mixed with absolute paths. - Source-Only Web Audits: Local web source folders are inspected for framework and dependency clues, including vulnerable React/Next.js combinations. - Source-Guided Web Exploits: Local source can drive live payloads for JSON length/type coercion and palindrome-style validation bugs while ignoring fake local flags when a spawned target is available. - Web Exploitation Playbooks: Browser-discovered forms can trigger archive upload, JSON/XML API, mass-assignment, XXE, JWT, and interesting-link follow-up checks. - Evidence-Led Web Artifact Triage: Web responses are inspected for header-disclosed artifacts such as X-Archived-Path and backup notes. The web agent can fetch those paths, decode certutil/PEM-style base64 blocks, reconstruct hidden SVG text from Krita-style archives, and render binary STL/OpenSCAD projections for manual flag reading. - Static-Source JWT Recovery: Live web targets are checked early for leaked JWT signing-key hints in HTML comments and static JavaScript. When a valid session token is present, the web agent can forge focused debug/admin claim variants and probe discovered chat/API endpoints without persisting raw secrets or forged tokens in artifacts. - Fast Live-Web Dispatch: If ask.py has already classified a prompt as a live web challenge with an explicit URL, the coordinator dispatches the first attempt directly to web_agent instead of waiting on LLM classification or routing. LLM-assisted recovery remains available after a failed specialist run. - HTB Code-Runner Playbooks: Web challenges exposing /run-style Python execution endpoints can submit compact solvers for coding/math tasks such as prime-product key recovery. - Remote ret2libc Pwn Playbook: Linux ELF pwn challenges with no PIE, NX enabled, a bundled libc.so.6, and a remote host:port can be exploited without local Linux execution. The pwn agent leaks puts, computes the libc base, builds a system("/bin/sh") chain, and retrieves common flag paths. - Hardware Logic Agent: Hardware/chip/circuit prompts can route to a specialist that combines challenge text, local files, images, and CSV tables to derive logic and decode output streams. Saleae .sal archives are inspected for analyzer metadata and decoded as UART 8N1 where applicable. - Godot Loader Reversing: Game-loader challenges can extract Godot PCK AES keys from Windows launchers, recover/decompile scripts with GDRE Tools, model GDScript obfuscation, and replay loader network requests to retrieve split flag material from headers and payload metadata. - Blockchain Specialist: Solidity folders can route to a Web3-backed blockchain agent that fetches HTB-style /connection_info, connects to the challenge RPC endpoint, executes deterministic contract exploits where applicable, and retrieves the remote flag. The normal ask.py path has been validated end to end against a spawned Survival-style smart-contract target. - Secure Coding Specialist: Secure-coding/source-remediation prompts route to a dedicated agent that uses editor-style APIs to inspect source, generate focused patches for recognized vulnerability patterns, save the updated file, and call the target's verification endpoint. The current playbook covers legacy flat-file user databases vulnerable to newline/pipe row injection. - Explicit Target Allowlisting: Remote challenge URLs, IP:port pairs, and connection-info endpoints must be approved through config/system_config.yaml or CTF_AGENTS_ALLOWED_NETWORKS; pasted challenge metadata cannot approve itself. The same policy is enforced across HTTP/browser tools, blockchain metadata fetches, raw crypto sockets, Docker readiness checks, nmap scans, and directory-discovery fallbacks. - Reduced Secret Exposure: Challenge-facing subprocesses run with a minimal environment by default so API keys and other host secrets are not inherited by LLM-generated scripts or untrusted challenge binaries unless a tool opts in. - Artifact Redaction By Default: Run reports, broker result messages, and the SQLite knowledge store redact sensitive keys before persistence. Browser cookies, Web Storage, and key-bearing generated scripts are not stored by default. - Opt-In Docker Challenge Runs: Local Docker web challenge folders can be built and launched when CTF_AGENTS_ALLOW_DOCKER=1 is set. - Live SSH Forensics: For authorized SSH-based forensics prompts, the forensics agent can inspect loader/preload state and shared-library hook indicators. Preload bypass searches require an explicit env opt-in.