CyberArena / app /generators /vulnerability_hunter.py
Hussien Haider
H
5c14256
Raw
History Blame Contribute Delete
70.3 kB
"""
vulnerability_hunter_generator.py
==================================
Generates fully-realized Vulnerability Hunter challenges for the CyberArena
Blue Team pool.
A "Vulnerability Hunter" challenge gives the trainee a small vulnerable code
sample and asks them to IDENTIFY the vulnerability (canonical key like
"sql-injection", "xss", "command-injection", ...) — not to fix the code.
This is the "defender triage" path of the Blue Team curriculum.
Architecture mirrors `code_fixing_generator.py`:
main.py watcher (polls vulnerability_hunter_challenges count)
|
v
refill_pool(team_role, count) --> ai_generate_challenge()
| |
| v
| AI generates vulnerable code
| |
| v
| ChallengeBuilder produces DB row
v
insert_to_db() <----- Challenge row
|
v
public.vulnerability_hunter_challenges
Difficulty uses 5 levels (Beginner / Easy / Medium / Hard / Expert), mapped
to Arabic values already in the platform: مبتدئ / سهل / متوسط / صعب / خبير.
Public API (used by main.py):
- POOL_TARGET = 5
- POOL_THRESHOLD = 2
- POOL_BATCH = 3
- get_pool_count(team_role) -> int
- async refill_pool(team_role, count) -> int
- async start_pool_watcher(team_role)
"""
from __future__ import annotations
import argparse
import asyncio
import json
import os
import random
import re
import sys
import time
import uuid
from typing import Optional
# Load .env early so the CLI works without manual export
try:
from dotenv import load_dotenv
_env_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), ".env")
if os.path.exists(_env_path):
load_dotenv(_env_path)
except ImportError:
pass
import httpx
# --------------------------------------------------------------------------- #
# 0. Constants
# --------------------------------------------------------------------------- #
ALLOWED_LANGUAGES = ("C++", "JAVA", "PYTHON", "JAVASCRIPT", "PHP", "RUST", "GO", "CSHARP", "RUBY", "CSHARP")
ALLOWED_DIFFICULTIES = ("مبتدئ", "سهل", "متوسط", "صعب", "خبير")
ALLOWED_TEAMS = ("blue",)
# Canonical vulnerability types — what the student must type. Order matters
# for the rotation. Deliberately broad: covers web, crypto, memory safety,
# concurrency, and language-specific pitfalls across C/C++/Java/Python/JS/PHP/
# Rust/Go/C#/Ruby.
VULNERABILITY_TYPES = [
# web / app
"sql-injection", "xss", "command-injection", "path-traversal",
"csrf", "ssrf", "xxe", "idor", "open-redirect",
"insecure-deserialization", "broken-authentication",
"security-misconfiguration", "sensitive-data-exposure",
"unrestricted-file-upload", "prototype-pollution",
"mass-assignment", "ssr-rce", "file-inclusion",
# crypto / secrets
"weak-hashing", "weak-randomness", "hardcoded-credentials",
"insecure-key-exchange", "ecb-mode", "padding-oracle",
# memory safety (C/C++/Rust)
"buffer-overflow", "use-after-free", "format-string",
"integer-overflow", "null-pointer-dereference",
"double-free", "off-by-one", "uninitialized-memory",
# concurrency / logic
"race-condition", "time-of-check-time-of-use",
"unsafe-reflection", "trust-boundary-violation",
]
# OWASP-family / class labels (used in hints + UI).
VULNERABILITY_CLASS = {
"sql-injection": "Injection",
"xss": "Injection (XSS)",
"command-injection": "Injection",
"path-traversal": "Path Traversal",
"csrf": "Cross-Site Request Forgery",
"ssrf": "Server-Side Request Forgery",
"xxe": "XML External Entity",
"xml-external-entity": "XML External Entity",
"idor": "Broken Access Control",
"insecure-direct-object-reference": "Broken Access Control",
"open-redirect": "Unvalidated Redirects",
"ssr-rce": "Server-Side Template Injection",
"file-inclusion": "File Inclusion",
"insecure-deserialization": "Insecure Deserialization",
"unsafe-deserialization": "Insecure Deserialization",
"broken-authentication": "Broken Authentication",
"security-misconfiguration": "Security Misconfiguration",
"sensitive-data-exposure": "Cryptographic Failures",
"unrestricted-file-upload": "File Upload",
"mass-assignment": "Mass Assignment",
"weak-hashing": "Cryptographic Failures",
"weak-randomness": "Cryptographic Failures",
"hardcoded-credentials": "Hardcoded Secrets",
"insecure-key-exchange": "Cryptographic Failures",
"ecb-mode": "Cryptographic Failures",
"padding-oracle": "Cryptographic Failures",
"buffer-overflow": "Memory Safety",
"use-after-free": "Memory Safety",
"format-string": "Memory Safety",
"integer-overflow": "Memory Safety",
"null-pointer-dereference": "Memory Safety",
"double-free": "Memory Safety",
"off-by-one": "Memory Safety",
"uninitialized-memory": "Memory Safety",
"race-condition": "Concurrency",
"time-of-check-time-of-use": "Concurrency",
"unsafe-reflection": "Input Validation",
"trust-boundary-violation": "Input Validation",
"prototype-pollution": "Prototype Pollution",
}
# Module names for each vulnerability (matches the existing /paths/modules).
# Web flaws land under web-security; crypto under cryptography; memory-safety /
# concurrency / language-specific bugs under secure-coding (a generic catch-all
# module for non-web code-review exercises).
MODULE_BY_VULN = {
"sql-injection": "web-security",
"xss": "web-security",
"command-injection": "web-security",
"path-traversal": "web-security",
"csrf": "web-security",
"ssrf": "web-security",
"xxe": "web-security",
"xml-external-entity": "web-security",
"idor": "web-security",
"insecure-direct-object-reference": "web-security",
"open-redirect": "web-security",
"ssr-rce": "web-security",
"file-inclusion": "web-security",
"insecure-deserialization": "web-security",
"unsafe-deserialization": "web-security",
"broken-authentication": "web-security",
"security-misconfiguration": "web-security",
"sensitive-data-exposure": "web-security",
"unrestricted-file-upload": "web-security",
"mass-assignment": "web-security",
"prototype-pollution": "web-security",
"weak-hashing": "cryptography",
"weak-randomness": "cryptography",
"hardcoded-credentials": "secure-coding",
"insecure-key-exchange": "cryptography",
"ecb-mode": "cryptography",
"padding-oracle": "cryptography",
"buffer-overflow": "secure-coding",
"use-after-free": "secure-coding",
"format-string": "secure-coding",
"integer-overflow": "secure-coding",
"null-pointer-dereference": "secure-coding",
"double-free": "secure-coding",
"off-by-one": "secure-coding",
"uninitialized-memory": "secure-coding",
"race-condition": "secure-coding",
"time-of-check-time-of-use": "secure-coding",
"unsafe-reflection": "secure-coding",
"trust-boundary-violation": "secure-coding",
}
# Pick a sensible default language per vuln type (the AI is told the language,
# not asked to choose). This list is what rotates through the seed pool.
LANG_BY_VULN = {
"sql-injection": "PYTHON",
"xss": "JAVASCRIPT",
"command-injection": "PYTHON",
"path-traversal": "PYTHON",
"csrf": "JAVASCRIPT",
"ssrf": "PYTHON",
"xxe": "JAVA",
"xml-external-entity": "JAVA",
"idor": "JAVASCRIPT",
"insecure-direct-object-reference": "JAVASCRIPT",
"open-redirect": "JAVASCRIPT",
"ssr-rce": "JAVASCRIPT",
"file-inclusion": "PHP",
"insecure-deserialization": "PYTHON",
"unsafe-deserialization": "PYTHON",
"broken-authentication": "JAVASCRIPT",
"security-misconfiguration": "PYTHON",
"sensitive-data-exposure": "PYTHON",
"unrestricted-file-upload": "PHP",
"mass-assignment": "RUBY",
"prototype-pollution": "JAVASCRIPT",
"weak-hashing": "PYTHON",
"weak-randomness": "JAVA",
"hardcoded-credentials": "CSHARP",
"insecure-key-exchange": "PYTHON",
"ecb-mode": "PYTHON",
"padding-oracle": "PYTHON",
"buffer-overflow": "C++",
"use-after-free": "C++",
"format-string": "C++",
"integer-overflow": "C++",
"null-pointer-dereference": "JAVA",
"double-free": "C++",
"off-by-one": "C++",
"uninitialized-memory": "C++",
"race-condition": "GO",
"time-of-check-time-of-use": "C++",
"unsafe-reflection": "JAVA",
"trust-boundary-violation": "PYTHON",
}
# Difficulty-level human descriptions (used inside the system prompt so the
# AI actually scales the vulnerable code complexity).
DIFFICULTY_PROFILE = {
"مبتدئ": {
"rank": 1,
"label": "Beginner",
"code_lines": (20, 30),
"obviousness": "very high — the vulnerability is the only interesting thing in the file",
},
"سهل": {
"rank": 2,
"label": "Easy",
"code_lines": (25, 40),
"obviousness": "high — one obvious flaw, minimal surrounding code",
},
"متوسط": {
"rank": 3,
"label": "Medium",
"code_lines": (35, 55),
"obviousness": "medium — vulnerability is hidden among 2-3 decoy patterns, the trainee must pick the right one",
},
"صعب": {
"rank": 4,
"label": "Hard",
"code_lines": (50, 75),
"obviousness": "low — multiple plausible-looking patterns, the real flaw is non-obvious and requires careful reading",
},
"خبير": {
"rank": 5,
"label": "Expert",
"code_lines": (70, 110),
"obviousness": "very low — production-quality code with a subtle flaw; multiple red herrings; requires deep expertise to spot",
},
}
# Pool constants
# Target: 5 challenges always available, refill when drops below 3
POOL_TARGET = 5
POOL_THRESHOLD = 2
POOL_BATCH = 3
from app.core.config import (
SUPABASE_URL, SUPABASE_ANON_KEY,
CLOUDFLARE_API_TOKEN, CLOUDFLARE_ACCOUNT_ID, CLOUDFLARE_MODEL, CLOUDFLARE_URL,
GROQ_API_KEY, GROQ_MODEL, GROQ_API_URL,
NVIDIA_API_KEY, NVIDIA_MODEL, NVIDIA_URL,
MISTRAL_API_KEY, MISTRAL_MODEL, MISTRAL_API_URL,
)
CLOUDFLARE_MODEL_FALLBACKS = [
"@cf/qwen/qwen2.5-coder-32b-instruct",
"@cf/meta/llama-3.3-70b-instruct-fp8-fast",
"@cf/meta/llama-3.1-70b-instruct",
"@cf/mistralai/mistral-small-3.1-24b-instruct",
"@cf/openai/gpt-oss-120b",
"@cf/openai/gpt-oss-20b",
"@cf/meta/llama-3.1-8b-instruct",
]
TABLE_NAME = "vulnerability_hunter_challenges"
# Per-team backoff tracker
_AI_BACKOFF_UNTIL: dict[str, float] = {}
# Concurrency control
_POOL_LOCKS: dict[str, asyncio.Lock] = {}
_WATCHER_STARTED: set[str] = set()
def _get_pool_lock(team_role: str) -> asyncio.Lock:
if team_role not in _POOL_LOCKS:
_POOL_LOCKS[team_role] = asyncio.Lock()
return _POOL_LOCKS[team_role]
# --------------------------------------------------------------------------- #
# 1. Helpers (mirror code_fixing_generator.py)
# --------------------------------------------------------------------------- #
def supabase_headers(content_type: bool = False) -> dict:
headers = {
"apikey": SUPABASE_ANON_KEY,
"Authorization": f"Bearer {SUPABASE_ANON_KEY}",
}
if content_type:
headers["Content-Type"] = "application/json"
return headers
def _extract_string_value(text: str, key: str) -> Optional[str]:
pattern = rf'"{re.escape(key)}"\s*:\s*"((?:[^"\\]|\\.)*)"'
match = re.search(pattern, text, re.DOTALL)
if match:
value = match.group(1)
value = value.replace('\\n', '\n').replace('\\t', '\t').replace('\\"', '"').replace('\\\\', '\\')
return value
return None
def _find_matching_bracket(text: str, open_idx: int, open_ch: str, close_ch: str) -> Optional[int]:
depth = 0
i = open_idx
in_string = False
escape_next = False
while i < len(text):
c = text[i]
if escape_next:
escape_next = False
i += 1
continue
if c == '\\' and in_string:
escape_next = True
i += 1
continue
if c == '"':
in_string = not in_string
elif not in_string:
if c == open_ch:
depth += 1
elif c == close_ch:
depth -= 1
if depth == 0:
return i
i += 1
return None
def parse_json_safe(raw) -> dict:
if raw is None or raw == "":
raise ValueError("Empty response from model")
if isinstance(raw, dict):
return raw
if not isinstance(raw, str):
raise ValueError(f"Expected str or dict, got {type(raw).__name__}")
cleaned = raw.strip()
fence_match = re.search(r"```(?:json)?\s*([\s\S]*?)```", cleaned, re.IGNORECASE)
if fence_match:
cleaned = fence_match.group(1).strip()
start = cleaned.find("{")
if start == -1:
raise ValueError(f"No JSON object found. Raw: {cleaned[:200]}")
end = _find_matching_bracket(cleaned, start, "{", "}")
if end is None:
end = cleaned.rfind("}")
if end == -1 or end <= start:
raise ValueError(f"Unbalanced JSON braces. Raw: {cleaned[:200]}")
cleaned = cleaned[start:end + 1]
cleaned = re.sub(r"[\x00-\x08\x0B\x0C\x0E-\x1F\x7F]", "", cleaned)
try:
return json.loads(cleaned, strict=False)
except json.JSONDecodeError:
pass
fixed = cleaned
fixed = re.sub(r',\s*([}\]])', r'\1', fixed)
fixed = re.sub(r'(?<!")(\b[A-Za-z_][A-Za-z0-9_]*\b)(\s*:)', r'"\1"\2', fixed)
fixed = re.sub(r"'([^'\\]*(?:\\.[^'\\]*)*)'", lambda m: '"' + m.group(1).replace('"', '\\"') + '"', fixed)
try:
return json.loads(fixed, strict=False)
except json.JSONDecodeError:
pass
result: dict = {}
for key in ("title", "story", "task_outline", "vulnerable_code",
"vulnerability_description", "difficulty",
"vulnerability_type", "vulnerability_class"):
val = _extract_string_value(cleaned, key)
if val is not None:
result[key] = val
hints_key_idx = cleaned.find('"hints"')
if hints_key_idx != -1:
bracket_start = cleaned.find("[", hints_key_idx)
if bracket_start != -1:
bracket_end = _find_matching_bracket(cleaned, bracket_start, "[", "]")
if bracket_end is not None:
hints_str = cleaned[bracket_start:bracket_end + 1]
try:
result["hints"] = json.loads(hints_str, strict=False)
except json.JSONDecodeError:
hints_str = re.sub(r',\s*([}\]])', r'\1', hints_str)
try:
result["hints"] = json.loads(hints_str, strict=False)
except Exception:
result["hints"] = []
if not result:
preview = raw[:500].replace("\n", " ")
raise ValueError(f"Could not extract any fields from model output. Raw preview: {preview}")
return result
def _pick_rotation_slot(i: int) -> tuple[str, str, str]:
"""Return (language, vulnerability_type, difficulty) for slot i."""
vuln = VULNERABILITY_TYPES[i % len(VULNERABILITY_TYPES)]
lang = LANG_BY_VULN.get(vuln, "PYTHON")
diff = ALLOWED_DIFFICULTIES[i % len(ALLOWED_DIFFICULTIES)]
return lang, vuln, diff
# --------------------------------------------------------------------------- #
# 2. Curated Seeds — one per (difficulty, vuln) hot spot so the pool is
# usable the moment migrations are applied.
# --------------------------------------------------------------------------- #
def _build_seeds() -> list[dict]:
"""Return the curated seed challenges used as a safety net when AI is
down. Coverage is intentionally broad across BOTH web and non-web
vulnerabilities in C / C++ / Java / Python / JS / PHP / Go / Rust / C# /
Ruby, and across the 5 difficulty levels (3 seeds per level = 15 total).
Each entry MUST be a complete, runnable program with a single
intentionally-planted flaw matching `vulnerability_type` (canonical
kebab-case key). Titles are short, no prefix.
"""
seeds: list[dict] = [
# ===================== مبتدئ (Beginner) =====================
{
"language": "PYTHON",
"vulnerability_type": "sql-injection",
"vulnerability_class": "Injection",
"module": "web-security",
"title": "استعلام تسجيل دخول قابل للحقن",
"story": "اكتشف فريق الأمن في شركة تجارة إلكترونية أن صفحة تسجيل الدخول تتعرض لمحاولات اختراق متكررة. تم اعتراض دالة `login()` التي يستدعيها السيرفر وُجد أنها تكتب اسم المستخدم مباشرة في الاستعلام.",
"task_outline": "ادرس الكود وحدد الثغرة الأمنية الوحيدة الموجودة فيه. اكتب اسم الثغرة بالإنجليزية بالضبط كما هي متعارف عليها في معايير OWASP.",
"vulnerable_code": "import sqlite3\n\ndef login(username, password):\n conn = sqlite3.connect('shop.db')\n cur = conn.cursor()\n query = f\"SELECT * FROM users WHERE username = '{username}' AND password = '{password}'\"\n cur.execute(query)\n row = cur.fetchone()\n conn.close()\n return row is not None\n",
"vulnerability_description": "استخدام f-string في بناء استعلام SQL يسمح للمهاجم بحقن تعليمات SQL إضافية، مثل ' OR '1'='1 لتجاوز المصادقة.",
"difficulty": "مبتدئ",
"xp_reward": 100,
"hints": [
{"level": 1, "text": "انتبه لكيفية دمج متغيرات المستخدم داخل الاستعلام", "xp_cost": 10},
{"level": 2, "text": "الاستعلام يُبنى عبر f-string بدلًا من parameterized query", "xp_cost": 20},
{"level": 3, "text": "اسم الثغرة: sql-injection", "xp_cost": 40},
],
},
{
"language": "C++",
"vulnerability_type": "buffer-overflow",
"vulnerability_class": "Memory Safety",
"module": "secure-coding",
"title": "قراءة اسم في مخزن ثابت بحجم 32 بايت",
"story": "برنامج ترحيب بسيط يطلب من المستخدم اسمه ويطبعه. لاحظ المختبر الأمني أنه عند إدخال نص طويل يتعطل البرنامج.",
"task_outline": "ادرس الكود وحدد نوع الثغرة. اكتب اسمها بالإنجليزية بالضبط.",
"vulnerable_code": "#include <iostream>\n#include <cstring>\n\nvoid greet() {\n char name[32];\n std::cout << \"Enter your name: \";\n std::cin >> name;\n std::cout << \"Hello, \" << name << \"\\n\";\n}\n\nint main() {\n greet();\n return 0;\n}\n",
"vulnerability_description": "std::cin >> مع مصفوفة ثابتة لا يحد من طول الإدخال، فيسمح بفيض في الذاكرة (Stack Buffer Overflow).",
"difficulty": "مبتدئ",
"xp_reward": 100,
"hints": [
{"level": 1, "text": "قارن حجم المخزن المؤقت مع ما يقبله std::cin >>", "xp_cost": 10},
{"level": 2, "text": "std::cin >> لا يحد من طول النص المُدخل", "xp_cost": 20},
{"level": 3, "text": "اسم الثغرة: buffer-overflow", "xp_cost": 40},
],
},
{
"language": "PYTHON",
"vulnerability_type": "weak-hashing",
"vulnerability_class": "Cryptographic Failures",
"module": "cryptography",
"title": "تجزئة كلمات المرور بـ MD5",
"story": "نظام إدارة مستخدمين يحفظ كلمات المرور بعد تجزئتها باستخدام MD5. لاحظ الفريق الأمني أن المهاجم يستطيع كسر كلمات المرور بسرعة باستخدام rainbow tables.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "import hashlib\n\ndef hash_password(pwd: str) -> str:\n return hashlib.md5(pwd.encode()).hexdigest()\n\ndef store_user(username, password):\n return {\n 'user': username,\n 'pwd_hash': hash_password(password),\n }\n",
"vulnerability_description": "MD5 دالة مجزّأة ضعيفة، سريعة، وبدون salt. يسهل كسرها.",
"difficulty": "مبتدئ",
"xp_reward": 100,
"hints": [
{"level": 1, "text": "ما نوع دالة التجزئة المستخدمة؟ هل هي آمنة لكلمات المرور؟", "xp_cost": 10},
{"level": 2, "text": "MD5 دالة قديمة وسريعة بدون salt", "xp_cost": 20},
{"level": 3, "text": "اسم الثغرة: weak-hashing", "xp_cost": 40},
],
},
# ===================== سهل (Easy) =====================
{
"language": "JAVA",
"vulnerability_type": "null-pointer-dereference",
"vulnerability_class": "Memory Safety",
"module": "secure-coding",
"title": "وصول لعضو قبل فحص null",
"story": "كلاس `CustomerService` يستدعي `customer.getName()` في معالجة الطلبات. لاحظ فريق المراقبة انهيار التطبيق عند تسجيل دخول ضيف (guest customer = null).",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "public class CustomerService {\n\n public String greet(Customer customer) {\n return \"Welcome, \" + customer.getName();\n }\n\n public static void main(String[] args) {\n CustomerService svc = new CustomerService();\n System.out.println(svc.greet(null));\n }\n}\n",
"vulnerability_description": "استدعاء customer.getName() على null قبل التحقق يرمي NullPointerException وقد يُستخدم لرفض الخدمة.",
"difficulty": "سهل",
"xp_reward": 120,
"hints": [
{"level": 1, "text": "هل يوجد فحص customer == null قبل الوصول لعضو؟", "xp_cost": 10},
{"level": 2, "text": "الاستدعاء يقع قبل أي فحص", "xp_cost": 20},
{"level": 3, "text": "اسم الثغرة: null-pointer-dereference", "xp_cost": 40},
],
},
{
"language": "JAVASCRIPT",
"vulnerability_type": "prototype-pollution",
"vulnerability_class": "Prototype Pollution",
"module": "secure-coding",
"title": "دمج عميق لكائن JSON دون تصفية",
"story": "مكتبة داخلية توفر دالة merge لدمج كائنات JSON. اكتشف الفريق الأمني أن المهاجم يستطيع تعديل Object.prototype مما يؤثر على كل الكائنات في النظام.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "function deepMerge(target, source) {\n for (const key of Object.keys(source)) {\n const sv = source[key];\n if (sv && typeof sv === 'object' && !Array.isArray(sv)) {\n target[key] = deepMerge(target[key] || {}, sv);\n } else {\n target[key] = sv;\n }\n }\n return target;\n}\n",
"vulnerability_description": "عدم تصفية المفاتيح __proto__ / constructor / prototype يسمح بتلويث النموذج الأولي العام.",
"difficulty": "سهل",
"xp_reward": 120,
"hints": [
{"level": 1, "text": "ماذا لو كان key = '__proto__'؟", "xp_cost": 10},
{"level": 2, "text": "لا توجد قائمة مفاتيح ممنوعة قبل الدمج العميق", "xp_cost": 20},
{"level": 3, "text": "اسم الثغرة: prototype-pollution", "xp_cost": 40},
],
},
{
"language": "CSHARP",
"vulnerability_type": "hardcoded-credentials",
"vulnerability_class": "Hardcoded Secrets",
"module": "secure-coding",
"title": "بيانات اعتماد مكتوبة في الكود",
"story": "خدمة داخلية تتصل بقاعدة بيانات الشركة. في الفحص الروتيني اكتشف فريق المراجعة أن اسم المستخدم وكلمة السر مكتوبتان في الكود المصدري.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "using System.Data.SqlClient;\n\npublic class Database {\n private const string DB_USER = \"sa\";\n private const string DB_PASS = \"Adm1n!2024#\";\n private const string DB_HOST = \"db.internal.corp\";\n\n public SqlConnection Open() {\n var conn = new SqlConnection(\n $\"Server={DB_HOST};Database=app;User Id={DB_USER};Password={DB_PASS};\");\n conn.Open();\n return conn;\n }\n}\n",
"vulnerability_description": "الكتابة المباشرة لكلمات السر في الكود تُعرّضها لتسريب المصدر (git, decompile, logs).",
"difficulty": "سهل",
"xp_reward": 120,
"hints": [
{"level": 1, "text": "هل كلمة السر في متغير بيئي أو vault؟", "xp_cost": 10},
{"level": 2, "text": "القيم مكتوبة في const string داخل الكود", "xp_cost": 20},
{"level": 3, "text": "اسم الثغرة: hardcoded-credentials", "xp_cost": 40},
],
},
# ===================== متوسط (Medium) =====================
{
"language": "C++",
"vulnerability_type": "format-string",
"vulnerability_class": "Memory Safety",
"module": "secure-coding",
"title": "سجل طباعة يستخدم مدخلات المستخدم كصيغة",
"story": "تطبيق خادم يطبع رسائل السجل باستخدام printf(message) حيث message يأتي من المستخدم. اكتشف فريق الأمن أن المهاجم يستطيع قراءة ذاكرة البرنامج أو تنفيذ كود.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "#include <cstdio>\n#include <cstring>\n\nvoid log_user_input(const char* user_msg) {\n char buf[256];\n snprintf(buf, sizeof(buf), \"[USER] \");\n strncat(buf, user_msg, 200);\n printf(buf);\n}\n",
"vulnerability_description": "printf بدون سلسلة تنسيق '%s' يسمح للمهاجم بحقن محددات مثل %x أو %n لقراءة/كتابة الذاكرة.",
"difficulty": "متوسط",
"xp_reward": 150,
"hints": [
{"level": 1, "text": "هل سلسلة التنسيق محددة في printf؟", "xp_cost": 15},
{"level": 2, "text": "printf(buf) يفسّر أي % في buf كمحدد", "xp_cost": 30},
{"level": 3, "text": "اسم الثغرة: format-string", "xp_cost": 50},
],
},
{
"language": "GO",
"vulnerability_type": "race-condition",
"vulnerability_class": "Concurrency",
"module": "secure-coding",
"title": "عدّاد مشترك بدون مزامنة في Go",
"story": "خدمة تتلقى طلبات كثيرة في نفس الوقت، وكل handler يزيد عدادًا مشتركًا. يلاحظ فريق المراقبة أن العداد يعطي قيمًا أقل من المتوقع.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "package main\n\nimport (\n \"fmt\"\n \"net/http\"\n)\n\nvar counter int\n\nfunc handler(w http.ResponseWriter, r *http.Request) {\n counter++\n fmt.Fprintf(w, \"count=%d\\n\", counter)\n}\n\nfunc main() {\n http.HandleFunc(\"/hit\", handler)\n http.ListenAndServe(\":8080\", nil)\n}\n",
"vulnerability_description": "counter++ ليس ذريًا، القراءة-الزيادة-الكتابة تتداخل بين الـ goroutines وتُفقد التحديثات.",
"difficulty": "متوسط",
"xp_reward": 150,
"hints": [
{"level": 1, "text": "هل يتم استخدام mutex أو atomic؟", "xp_cost": 15},
{"level": 2, "text": "counter++ = LOAD + ADD + STORE، تتخللها goroutines أخرى", "xp_cost": 30},
{"level": 3, "text": "اسم الثغرة: race-condition", "xp_cost": 50},
],
},
{
"language": "PYTHON",
"vulnerability_type": "insecure-deserialization",
"vulnerability_class": "Insecure Deserialization",
"module": "secure-coding",
"title": "تحميل pickle من ملف يتحكم به المستخدم",
"story": "تطبيق يخزن الجلسات في ملفات على القرص، وعند كل طلب يتم تحميلها عبر pickle.load. المهاجم يستطيع رفع ملف جلسة مزيف لتنفيذ كود على السيرفر.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "import pickle\nimport os\n\ndef load_session(session_id):\n path = os.path.join('/tmp/sessions', session_id)\n with open(path, 'rb') as f:\n return pickle.load(f)\n",
"vulnerability_description": "pickle.load على بيانات يتحكم بها المهاجم ينفذ كود Python تعسفي عند الإلغاء عبر __reduce__.",
"difficulty": "متوسط",
"xp_reward": 150,
"hints": [
{"level": 1, "text": "هل تثق بمحتوى الملف الذي يحمّله pickle.load؟", "xp_cost": 15},
{"level": 2, "text": "pickle يستدعي __reduce__ الذي قد ينفذ أوامر", "xp_cost": 30},
{"level": 3, "text": "اسم الثغرة: insecure-deserialization", "xp_cost": 50},
],
},
# ===================== صعب (Hard) =====================
{
"language": "JAVA",
"vulnerability_type": "unsafe-reflection",
"vulnerability_class": "Input Validation",
"module": "secure-coding",
"title": "استدعاء Class.forName على مدخلات المستخدم",
"story": "نظام إضافة إضافات (plugins) يستدعي Class.forName باسم يحدده المستخدم، ثم يستدعي newInstance. لاحظ فريق الأمن أن المهاجم يستطيع تشغيل أي كلاس على الـ classpath.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "public class PluginLoader {\n\n public static Object load(String className) throws Exception {\n Class<?> cls = Class.forName(className);\n return cls.getDeclaredConstructor().newInstance();\n }\n\n public static void main(String[] args) throws Exception {\n Object plugin = load(args[0]);\n System.out.println(\"loaded: \" + plugin);\n }\n}\n",
"vulnerability_description": "Class.forName على مدخلات المستخدم يسمح بتحميل أي كلاس متاح (RCE عبر Runtime.exec مثلًا).",
"difficulty": "صعب",
"xp_reward": 200,
"hints": [
{"level": 1, "text": "هل className يُتحقق منه ضد whitelist؟", "xp_cost": 20},
{"level": 2, "text": "Class.forName يقبل أي اسم متاح في الـ classpath", "xp_cost": 40},
{"level": 3, "text": "اسم الثغرة: unsafe-reflection", "xp_cost": 70},
],
},
{
"language": "RUBY",
"vulnerability_type": "mass-assignment",
"vulnerability_class": "Mass Assignment",
"module": "secure-coding",
"title": "تحديث نموذج Ruby on Rails مع جميع المعاملات",
"story": "نموذج User في تطبيق Rails يقبل كل المعاملات من params ويحدّثها. اكتشف فريق الأمن أن المهاجم يرسل admin=true في JSON ويصبح مديرًا.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "class UsersController < ApplicationController\n def update\n user = User.find(params[:id])\n user.update(params[:user])\n render json: user\n end\nend\n",
"vulnerability_description": "تمرير params[:user] مباشرةً إلى update يسمح للمهاجم بضبط أعمدة حساسة (is_admin, role) لم يكن من المفترض أن يعدّلها.",
"difficulty": "صعب",
"xp_reward": 200,
"hints": [
{"level": 1, "text": "هل يوجد strong parameters أو attr_accessible؟", "xp_cost": 20},
{"level": 2, "text": "params[:user] قد يحوي admin: true", "xp_cost": 40},
{"level": 3, "text": "اسم الثغرة: mass-assignment", "xp_cost": 70},
],
},
{
"language": "C++",
"vulnerability_type": "double-free",
"vulnerability_class": "Memory Safety",
"module": "secure-coding",
"title": "تحرير نفس المخزن مرتين في مسار الخطأ",
"story": "دالة تعالج مدخلات في مخزن مؤقت. في حالة فشل المعالجة، تُستدعى free(buf) مرتين في فروع شرطية متقاربة، مما يتسبب في تعطل البرنامج أو تنفيذ كود.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "#include <cstdlib>\n#include <cstring>\n\nvoid process(const char* data, int len) {\n char* buf = (char*)malloc(len);\n memcpy(buf, data, len);\n if (len < 0) {\n free(buf);\n return;\n }\n if (len == 0) {\n free(buf);\n return;\n }\n free(buf);\n}\n",
"vulnerability_description": "تحرير الذاكرة مرتين (double-free) يسمح للمهاجم بالسيطرة على allocator وتنفيذ كود.",
"difficulty": "صعب",
"xp_reward": 200,
"hints": [
{"level": 1, "text": "هل buf = nullptr بعد كل free؟", "xp_cost": 20},
{"level": 2, "text": "إذا مرّ المسار بفروع متتالية قد يُحرر buf مرتين", "xp_cost": 40},
{"level": 3, "text": "اسم الثغرة: double-free", "xp_cost": 70},
],
},
# ===================== خبير (Expert) =====================
{
"language": "C++",
"vulnerability_type": "integer-overflow",
"vulnerability_class": "Memory Safety",
"module": "secure-coding",
"title": "حساب حجم المخزن عبر ضرب int*sizeof",
"story": "دالة تحجز مخزنًا بناءً على ضرب حجم العنصر في عدد العناصر. في حالة تجاوز حاصل الضرب للحد الأقصى، يُحجز مخزن صغير جدًا ويسبب كتابة خارج الحدود.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "#include <cstdlib>\n#include <cstring>\n\nvoid copy(int count, const char* data) {\n size_t total = count * sizeof(int);\n char* buf = (char*)malloc(total);\n memcpy(buf, data, total);\n}\n",
"vulnerability_description": "count * sizeof(int) قد يفيض إذا كان count كبيرًا، فيُحجز مخزن أصغر من المتوقع ثم تُنسخ البيانات كاملةً إلى خارج حدوده.",
"difficulty": "خبير",
"xp_reward": 250,
"hints": [
{"level": 1, "text": "ماذا لو كان count قريبًا من INT_MAX؟", "xp_cost": 30},
{"level": 2, "text": "الضرب count * sizeof(int) يتجاوز size_t؟ كلاهما", "xp_cost": 60},
{"level": 3, "text": "اسم الثغرة: integer-overflow", "xp_cost": 100},
],
},
{
"language": "PYTHON",
"vulnerability_type": "time-of-check-time-of-use",
"vulnerability_class": "Concurrency",
"module": "secure-coding",
"title": "سباق بين فحص الصلاحية والاستخدام",
"story": "خدمة تحقّق أن المستخدم يملك رصيدًا كافيًا قبل تنفيذ السحب. اكتشف فريق الأمن أن المهاجم يرسل طلبين متزامنين بنفس المعرّف فيخصم الرصيد مرتين.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "balances = {}\n\ndef withdraw(user, amount):\n cur = balances.get(user, 0)\n if cur >= amount:\n balances[user] = cur - amount\n return True\n return False\n",
"vulnerability_description": "الفحص (>=) والاستخدام (-) ليسا ذريين. بين الخطوتين يقرأ طلب آخر نفس الرصيد ويخصمه أيضًا (TOCTOU).",
"difficulty": "خبير",
"xp_reward": 250,
"hints": [
{"level": 1, "text": "هل الخطوتان داخل عملية ذرية (قفل/معاملة)؟", "xp_cost": 30},
{"level": 2, "text": "هناك فجوة زمنية بين القراءة والكتابة", "xp_cost": 60},
{"level": 3, "text": "اسم الثغرة: time-of-check-time-of-use", "xp_cost": 100},
],
},
{
"language": "JAVA",
"vulnerability_type": "weak-randomness",
"vulnerability_class": "Cryptographic Failures",
"module": "cryptography",
"title": "توليد توكن باستخدام java.util.Random",
"story": "خدمة تولّد توكن استعادة كلمة المرور عبر java.util.Random ثم تحوّله إلى Base64. المهاجم يلاحظ أن قيم التوكنات متسلسلة ويسطو على حسابات المستخدمين.",
"task_outline": "ادرس الكود وحدد نوع الثغرة.",
"vulnerable_code": "import java.util.Random;\n\npublic class TokenGen {\n private final Random rng = new Random();\n\n public String nextToken() {\n long v = rng.nextLong();\n return Long.toHexString(v);\n }\n\n public static void main(String[] args) {\n TokenGen g = new TokenGen();\n for (int i = 0; i < 5; i++) {\n System.out.println(g.nextToken());\n }\n }\n}\n",
"vulnerability_description": "java.util.Random يستخدم LCG بذرة مكشوفة، فيمكن إعادة بناء البذرة من عيّنات متتالية وتوقع التوكنات القادمة.",
"difficulty": "خبير",
"xp_reward": 250,
"hints": [
{"level": 1, "text": "هل Random آمن للاستخدامات الأمنية؟", "xp_cost": 30},
{"level": 2, "text": "java.util.Random = LCG، قابل للتوقع", "xp_cost": 60},
{"level": 3, "text": "اسم الثغرة: weak-randomness", "xp_cost": 100},
],
},
]
return seeds
# --------------------------------------------------------------------------- #
# 3. AI Generation
# --------------------------------------------------------------------------- #
SYSTEM_PROMPT_TEMPLATE = """You are a cybersecurity expert who generates realistic vulnerable code samples for a "Vulnerability Hunter" training game.
=========================================================
GAME MECHANICS — READ THIS FIRST AND OBEY STRICTLY
=========================================================
This is NOT a "code fixing" game. There is no code editor for the trainee.
There is no patched/fixed version, no "modify this line", no "rewrite this function".
The trainee's ONLY interaction with the challenge is:
1. They SEE a code snippet on screen (read-only).
2. They TYPE the canonical vulnerability name (e.g. "sql-injection", "buffer-overflow")
into a single text input field.
3. They press "Submit". The backend compares their typed string against the
canonical key and reports right/wrong.
That is the entire game. The trainee never edits, never compiles, never runs,
never refactors, never "fixes" the code. The code is a static exhibit; the only
task is to name the flaw.
=========================================================
WHAT YOU MUST PRODUCE
=========================================================
- A complete, syntactically correct, runnable program in {language} of between
{code_min} and {code_max} non-blank lines, with exactly ONE primary
vulnerability of type "{vuln_type}" (OWASP class: {vuln_class}).
- The code is a FORENSIC EVIDENCE — frozen, not to be modified.
- An Arabic task_outline that ONLY tells the trainee: "read the code, identify
the vulnerability, and type its canonical English name in the input field".
DO NOT write: "fix the code", "patch the bug", "rewrite the function",
"edit line N", "correct the error", or anything that implies editing.
The target canonical vulnerability is: {vuln_type}
The target OWASP family / class is: {vuln_class}
The target language is: {language}
The target difficulty is: {difficulty} ({difficulty_label}) — {difficulty_obviousness}
DIFFICULTY SCALING:
- Code length MUST be between {code_min} and {code_max} lines (non-blank, non-comment).
- For مبتدئ / سهل: vulnerability is the obvious focal point. Minimal surrounding code.
- For متوسط: 1 vulnerability hidden among 2-3 plausible-looking but harmless patterns.
- For صعب / خبير: realistic production code. Multiple imports, helper functions,
configuration, docstrings. The real flaw is subtle and requires careful reading;
multiple red herrings.
ABSOLUTE RULES — VIOLATIONS WILL BE REJECTED:
1. Respond with ONLY valid JSON. No markdown, no code fences, no explanation
before or after.
2. The "vulnerable_code" field MUST be a complete, syntactically correct,
fully-implemented program of EXACTLY between {code_min} and {code_max}
non-blank lines. All imports, function definitions, and a main entry point
must be present. No truncation, no "...", no half-written functions, no
"# TODO" placeholders, no "see fix below" comments.
3. There must be exactly ONE primary vulnerability of type "{vuln_type}".
No other major issues.
4. "vulnerability_description" must explain the specific flaw and how an
attacker could exploit it.
5. "task_outline" MUST be a single sentence in Arabic, in EXACTLY this shape
(translated, paraphrased freely but keeping the meaning):
"ادرس الكود التالي جيداً، حدّد الثغرة الأمنية الوحيدة الموجودة فيه،
ثم اكتب اسمها بالإنجليزية (مفتاح قصير) في حقل الإجابة."
DO NOT mention: إصلاح، أصلح، صحّح، عدّل، طبّق، اكتب الكود الآمن، ضع، غيّر،
rewrite, fix, patch, edit, modify, correct. These are FORBIDDEN words in
task_outline. The trainee's job is IDENTIFICATION, not REMEDIATION.
6. "vulnerability_type" MUST be the canonical key "{vuln_type}" (English,
kebab-case). "vulnerability_class" MUST be "{vuln_class}".
7. ALL user-facing text fields (title, story, task_outline,
vulnerability_description, hints) MUST be written in Arabic ONLY. No
English, no Chinese, no other languages. The "vulnerable_code" field
contains code in {language} (which is correct), but all comments inside
the code should be in English (//, #, etc.) since this is conventional.
8. "difficulty" must be EXACTLY one of these Arabic words:
"مبتدئ", "سهل", "متوسط", "صعب", or "خبير".
9. "hints" must be a JSON array of exactly 3 objects:
[{{"level": 1, "text": "تلميح أول غامض", "xp_cost": 10}},
{{"level": 2, "text": "تلميح ثانٍ أكثر تحديداً", "xp_cost": 20}},
{{"level": 3, "text": "تلميح ثالث يكشف اسم الثغرة", "xp_cost": 40}}]
The first hint should be vague, the second more specific, the third
should reveal the canonical vulnerability name.
UNIQUENESS RULES — ZERO TOLERANCE FOR DUPLICATES:
- Every challenge MUST be completely unique in title, story, code structure, and variable names.
- NEVER reuse the same application domain (e.g. if previous was a "bank API", make this one a "hospital system", "IoT firmware", "game server", "scientific computing", etc.).
- NEVER reuse variable names, function names, class names, or file names from common templates. Use creative, context-specific naming.
- The vulnerable_code MUST use a different programming pattern, different data structures, and different business logic each time.
- The title and story MUST describe a completely different scenario — not just a paraphrase.
- Vary the code style: different imports, different module structure, different comments.
- DONT GENERATE ANY OF THE FOLLOWING TITLES (THEY ARE ALREADY IN THE DATABASE):
{blacklisted_titles}
Return ONLY this exact JSON structure (no other text):
{{"title": "عنوان التحدي بالعربية", "story": "قصة التحدي بالعربية", "task_outline": "ادرس الكود التالي وحدد الثغرة الأمنية الوحيدة الموجودة فيه ثم اكتب اسمها بالإنجليزية (مفتاح قصير) في حقل الإجابة.", "vulnerable_code": "الكود الكامل المصاب بالثغرة", "vulnerability_type": "{vuln_type}", "vulnerability_class": "{vuln_class}", "vulnerability_description": "شرح الثغرة بالعربية", "difficulty": "{difficulty}", "hints": [{{"level": 1, "text": "...", "xp_cost": 10}}, {{"level": 2, "text": "...", "xp_cost": 20}}, {{"level": 3, "text": "...", "xp_cost": 40}}]}}
"""
async def _post_with_json_fallback(client, url: str, payload: dict, headers: dict, provider_name: str) -> Optional[httpx.Response]:
resp = await client.post(url, json=payload, headers=headers)
if resp.status_code in (400, 404, 422) and "response_format" in payload:
print(f" [{provider_name}] response_format not supported ({resp.status_code}), retrying without...")
payload2 = {k: v for k, v in payload.items() if k != "response_format"}
resp = await client.post(url, json=payload2, headers=headers)
return resp
async def _try_mistral(prompt: str, system: str, model: str) -> Optional[str]:
"""Try Mistral API (mistral-large-latest by default).
Per the user's request, Mistral is the PRIMARY tier for challenge
generation — it runs first in the orchestrator. Only on a hard
failure (429 / timeout / 5xx / parse) do we fall through to
Cloudflare, then Groq, then NVIDIA/DeepSeek, then seed.
"""
if not MISTRAL_API_KEY:
return None
import httpx as _httpx
try:
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {MISTRAL_API_KEY}",
}
messages = []
if system:
messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
payload = {
"model": model or MISTRAL_MODEL,
"messages": messages,
"temperature": 0.7,
"max_tokens": 4096,
"response_format": {"type": "json_object"},
}
async with _httpx.AsyncClient(timeout=30) as client:
resp = await _post_with_json_fallback(client, MISTRAL_API_URL, payload, headers, "mistral")
if resp.status_code == 200:
data = resp.json()
content = data.get("choices", [{}])[0].get("message", {}).get("content")
if content:
return content
print(f" [mistral] 200 but empty content: {str(data)[:200]}")
return None
if resp.status_code == 429:
print(f" [mistral] 429 rate-limited")
else:
print(f" [mistral] error {resp.status_code}: {resp.text[:200]}")
return None
except Exception as e:
print(f" [mistral] exception: {type(e).__name__}: {e}")
return None
async def _try_cloudflare(prompt: str, system: str) -> Optional[str]:
if not (CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID):
return None
import httpx as _httpx
model_candidates = [CLOUDFLARE_MODEL] + [m for m in CLOUDFLARE_MODEL_FALLBACKS if m != CLOUDFLARE_MODEL]
messages = []
if system:
messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
payload_base = {"messages": messages, "temperature": 0.7, "max_tokens": 4096}
for model_name in model_candidates:
url = f"https://api.cloudflare.com/client/v4/accounts/{CLOUDFLARE_ACCOUNT_ID}/ai/run/{model_name}"
headers = {"Content-Type": "application/json", "Authorization": f"Bearer {CLOUDFLARE_API_TOKEN}"}
payload = {**payload_base, "response_format": {"type": "json_object"}}
try:
async with _httpx.AsyncClient(timeout=30) as client:
resp = await _post_with_json_fallback(client, url, payload, headers, f"cloudflare/{model_name.split('/')[-1]}")
if resp.status_code == 200:
result = resp.json()
if result.get("success") and result.get("result", {}).get("response"):
return result["result"]["response"]
print(f" [cloudflare/{model_name.split('/')[-1]}] 200 but no response field")
continue
if resp.status_code == 429:
print(f" [cloudflare/{model_name.split('/')[-1]}] 429 rate-limited")
continue
print(f" [cloudflare/{model_name.split('/')[-1]}] error {resp.status_code}: {resp.text[:120]}")
except Exception as e:
print(f" [cloudflare/{model_name.split('/')[-1]}] exception: {type(e).__name__}: {e}")
continue
return None
async def _try_groq(prompt: str, system: str, model: str) -> Optional[str]:
if not GROQ_API_KEY:
return None
import httpx as _httpx
try:
headers = {"Content-Type": "application/json", "Authorization": f"Bearer {GROQ_API_KEY}"}
messages = []
if system:
messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
payload = {
"model": model or GROQ_MODEL,
"messages": messages,
"temperature": 0.7,
"max_tokens": 4096,
"response_format": {"type": "json_object"},
}
async with _httpx.AsyncClient(timeout=90) as client:
resp = await _post_with_json_fallback(client, GROQ_API_URL, payload, headers, "groq")
if resp.status_code == 200:
data = resp.json()
content = data.get("choices", [{}])[0].get("message", {}).get("content")
if content:
return content
print(f" [groq] 200 but empty content: {str(data)[:200]}")
return None
if resp.status_code == 429:
print(f" [groq] 429 rate-limited")
else:
print(f" [groq] error {resp.status_code}: {resp.text[:200]}")
return None
except Exception as e:
print(f" [groq] exception: {type(e).__name__}: {e}")
return None
async def _try_nvidia(prompt: str, system: str, model: str) -> Optional[str]:
if not NVIDIA_API_KEY:
return None
import httpx as _httpx
try:
headers = {"Content-Type": "application/json", "Authorization": f"Bearer {NVIDIA_API_KEY}"}
messages = []
if system:
messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
payload = {
"model": model or NVIDIA_MODEL,
"messages": messages,
"temperature": 0.7,
"max_tokens": 4096,
"response_format": {"type": "json_object"},
}
async with _httpx.AsyncClient(timeout=90) as client:
resp = await _post_with_json_fallback(client, NVIDIA_API_URL, payload, headers, "deepseek")
if resp.status_code == 200:
data = resp.json()
content = data.get("choices", [{}])[0].get("message", {}).get("content")
if content:
return content
print(f" [deepseek] 200 but empty content: {str(data)[:200]}")
return None
if resp.status_code == 429:
print(f" [deepseek] 429 rate-limited")
else:
print(f" [deepseek] error {resp.status_code}: {resp.text[:200]}")
return None
except Exception as e:
print(f" [deepseek] exception: {type(e).__name__}: {e}")
return None
async def _call_ai(prompt: str, system: str = "", model: str = "") -> Optional[str]:
print(f" [ai] Trying Mistral...")
r = await _try_mistral(prompt, system, model)
if r:
return r
print(f" [ai] Trying Cloudflare...")
r = await _try_cloudflare(prompt, system)
if r:
return r
print(f" [ai] Trying Groq...")
r = await _try_groq(prompt, system, model)
if r:
return r
print(f" [ai] Trying NVIDIA/DeepSeek...")
r = await _try_nvidia(prompt, system, model)
if r:
return r
print(f" [ai] All 4 providers failed")
return None
async def ai_generate_challenge(language: str, vuln_type: str, difficulty: str, blacklisted_titles: str = "") -> Optional[dict]:
vuln_class = VULNERABILITY_CLASS.get(vuln_type, "Security Misconfiguration")
diff_profile = DIFFICULTY_PROFILE.get(difficulty, DIFFICULTY_PROFILE["متوسط"])
system = SYSTEM_PROMPT_TEMPLATE.format(
vuln_type=vuln_type,
vuln_class=vuln_class,
language=language,
difficulty=difficulty,
difficulty_label=diff_profile["label"],
difficulty_obviousness=diff_profile["obviousness"],
code_min=diff_profile["code_lines"][0],
code_max=diff_profile["code_lines"][1],
blacklisted_titles=blacklisted_titles or "لا يوجد عناوين سابقة حالياً.",
)
base_prompt = f"ولّد كود {language} بطول {diff_profile['code_lines'][0]}-{diff_profile['code_lines'][1]} سطر يحتوي على ثغرة {vuln_type} (OWASP: {vuln_class}) وفق الهيكل المطلوب. مستوى الصعوبة: {diff_profile['label']}."
# Round 1
raw = await _call_ai(base_prompt, system)
data = _try_parse(raw)
built = _validate_and_build(data, language, vuln_type, difficulty) if data else None
if built:
return built
# Round 2
retry_prompt = (
f"CRITICAL: Return ONLY valid JSON, no markdown, no code fences, no comments. "
f"Generate a complete {diff_profile['code_lines'][0]}-{diff_profile['code_lines'][1]} line {language} program "
f"with a {vuln_type} vulnerability (class: {vuln_class}). Difficulty: {diff_profile['label']}. "
f"Return exactly this structure with Arabic in all string fields:\n"
f'{{"title":"...","story":"...","task_outline":"...","vulnerable_code":"...","vulnerability_type":"{vuln_type}","vulnerability_class":"{vuln_class}","vulnerability_description":"...","difficulty":"{difficulty}","hints":[{{"level":1,"text":"...","xp_cost":10}},{{"level":2,"text":"...","xp_cost":20}},{{"level":3,"text":"...","xp_cost":40}}]}}'
)
print(f" [ai] Round 2: retrying with stricter prompt")
raw2 = await _call_ai(retry_prompt, system)
data2 = _try_parse(raw2)
built2 = _validate_and_build(data2, language, vuln_type, difficulty) if data2 else None
if built2:
return built2
# Round 3
minimal_prompt = (
f"Output JSON only. language={language}, vuln={vuln_type}, class={vuln_class}, "
f"difficulty={difficulty}. Required keys: title, story, task_outline, vulnerable_code "
f"(complete {diff_profile['code_lines'][0]}-{diff_profile['code_lines'][1]} line program), "
f"vulnerability_type, vulnerability_class, vulnerability_description, difficulty, "
f"hints (array of 3 objects with level,text,xp_cost). No markdown. No commentary."
)
print(f" [ai] Round 3: minimal retry")
raw3 = await _call_ai(minimal_prompt, system)
data3 = _try_parse(raw3)
built3 = _validate_and_build(data3, language, vuln_type, difficulty) if data3 else None
if built3:
return built3
print(f" [ai] All rounds exhausted for {language}/{vuln_type}/{difficulty}")
return None
def _try_parse(raw) -> Optional[dict]:
if not raw:
return None
if isinstance(raw, dict):
return raw
try:
return parse_json_safe(raw)
except Exception as e:
print(f" [ai] parse_json_safe failed: {e}")
return None
# Words that betray the AI's "code-fixing" bias. If any of these appear in
# task_outline, the challenge is REJECTED — the trainee's job is identification,
# not remediation. The model has to learn that this is a triage game.
FORBIDDEN_TASK_WORDS_AR = [
"أصلح", "إصلاح", "صحّح", "تصحيح", "عدّل", "تعديل", "طبّق", "تطبيق",
"اكتب الكود", "ضع الكود", "غيّر", "غيّر الكود", "استبدل", "أعد كتابة",
"برمج", "أنشئ دالة", "أنشئ كود", "أنشئ الإصدار", "اكتب دالة", "اكتب إصدار",
"طبّق", "نفّذ", "بادر", "احذف", "أزل", "تخلص من",
]
FORBIDDEN_TASK_WORDS_EN = [
"fix the", "fix this", "patch the", "patch this", "rewrite the",
"rewrite this", "modify the", "modify this", "edit the", "edit this",
"correct the", "correct this", "replace the", "replace this",
"refactor the", "refactor this", "implement a fix", "write a fix",
"write a patch", "secure version", "fixed version", "fixed code",
"update the code", "change the code", "apply a fix",
]
def _strip_fix_bias(text: str) -> str:
"""Post-process the AI's task_outline to neutralize any 'fix the code' bias.
The game is identification-only. If the AI slipped in repair instructions,
we replace the whole sentence with the canonical identification directive.
"""
if not text:
return text
low = text.lower()
for w in FORBIDDEN_TASK_WORDS_AR + FORBIDDEN_TASK_WORDS_EN:
if w.lower() in low:
return ("ادرس الكود التالي جيداً، حدّد الثغرة الأمنية الوحيدة الموجودة فيه، "
"ثم اكتب اسمها بالإنجليزية (مفتاح قصير) في حقل الإجابة.")
return text
def _validate_and_build(data: dict, language: str, vuln_type: str, difficulty: str) -> Optional[dict]:
required = ["title", "story", "task_outline", "vulnerable_code",
"vulnerability_description", "difficulty",
"vulnerability_type", "vulnerability_class"]
for f in required:
if not data.get(f):
print(f" [ai] Missing field: {f}")
return None
vcode = data.get("vulnerable_code", "")
line_count = len([l for l in vcode.split("\n") if l.strip() and not l.strip().startswith(("#", "//"))])
if line_count < 12:
print(f" [ai] Rejected: vulnerable_code too short ({line_count} effective lines)")
return None
# Acceptable diffs are only the 5 levels
if data["difficulty"] not in ALLOWED_DIFFICULTIES:
data["difficulty"] = difficulty if difficulty in ALLOWED_DIFFICULTIES else "متوسط"
# Reject mismatched vuln types
if data.get("vulnerability_type") and data["vulnerability_type"] != vuln_type:
# Trust the request slot — overwrite
data["vulnerability_type"] = vuln_type
# Strip any "fix the code" bias from the AI's task_outline
raw_task = data.get("task_outline", "")
sanitized_task = _strip_fix_bias(raw_task)
if sanitized_task != raw_task:
print(f" [ai] task_outline had fix-bias; replaced with canonical identification directive")
hints = data.get("hints") or []
if isinstance(hints, str):
try:
hints = json.loads(hints)
except Exception:
hints = []
if not isinstance(hints, list):
hints = []
xp = data.get("xp_reward") or {"خبير": 250, "صعب": 200, "متوسط": 150, "سهل": 120, "مبتدئ": 100}.get(data["difficulty"], 150)
return {
"team_role": "blue",
"language": language,
"module": MODULE_BY_VULN.get(vuln_type, "web-security"),
"title": data["title"],
"story": data["story"],
"task_outline": sanitized_task,
"vulnerable_code": vcode,
"vulnerability_type": vuln_type,
"vulnerability_class": VULNERABILITY_CLASS.get(vuln_type, data.get("vulnerability_class") or "Security Misconfiguration"),
"vulnerability_description": data.get("vulnerability_description", ""),
"hints": hints,
"difficulty": data["difficulty"],
"xp_reward": int(xp),
}
# --------------------------------------------------------------------------- #
# 4. DB Operations
# --------------------------------------------------------------------------- #
async def get_pool_count(team_role: str) -> int:
if not SUPABASE_ANON_KEY or not SUPABASE_URL:
return 0
url = f"{SUPABASE_URL}/rest/v1/{TABLE_NAME}?select=id&team_role=eq.{team_role}"
try:
async with httpx.AsyncClient() as client:
resp = await client.get(url, headers=supabase_headers())
if resp.status_code == 200:
return len(resp.json())
except Exception as e:
print(f"[vuln-hunter] Error checking pool count: {e}")
return 0
async def _insert_to_db(row: dict) -> bool:
if not SUPABASE_ANON_KEY or not SUPABASE_URL:
return False
from app.core.config import normalize_row_module
row = normalize_row_module(TABLE_NAME, row)
from app.services.insert_guard import atomic_insert
from app.services.dedup import is_duplicate_vuln_hunter
team_role = row.get("team_role", "blue")
return await atomic_insert(
table=TABLE_NAME,
team_role=team_role,
row=row,
dedup_func=is_duplicate_vuln_hunter,
dedup_args=[
row.get("title", ""),
row.get("vulnerable_code", ""),
row.get("task_outline", ""),
],
dedup_kwargs={"role_filter": team_role},
)
async def _delete_challenge(challenge_id: str):
if not SUPABASE_ANON_KEY or not SUPABASE_URL:
return
url = f"{SUPABASE_URL}/rest/v1/{TABLE_NAME}?id=eq.{challenge_id}"
try:
async with httpx.AsyncClient() as client:
await client.delete(url, headers=supabase_headers())
except Exception as e:
print(f" Delete error: {e}")
# --------------------------------------------------------------------------- #
# 5. Pool Refill
# --------------------------------------------------------------------------- #
async def _refill_with_seeds_only(team_role: str, count: int) -> int:
"""Insert `count` curated seeds (no AI) — used for pre-warm."""
seeds = _build_seeds()
random.shuffle(seeds)
inserted = 0
for seed in seeds[:count]:
if await _insert_to_db(seed):
inserted += 1
return inserted
async def _refill_pool_inner(team_role: str, count: int) -> int:
"""Unlocked inner — caller must hold _get_pool_lock(team_role)."""
if team_role not in ALLOWED_TEAMS:
return 0
until = _AI_BACKOFF_UNTIL.get(team_role, 0)
if time.time() < until:
print(f" [{team_role}] Backoff active until {until:.0f}, using seeds only")
return await _refill_with_seeds_only(team_role, count)
base_count = await get_pool_count(team_role)
needed = max(0, POOL_TARGET - base_count)
target = min(count, needed)
if target <= 0:
return 0
recent_titles = []
try:
from app.services.dedup import fetch_existing_titles
recent_titles = fetch_existing_titles("vulnerability_hunter_challenges", team_role, limit=30)
except Exception as e:
print(f"[vuln-hunter] Failed to fetch recent titles: {e}")
blacklisted_titles_str = "\n".join([f"- {t}" for t in recent_titles]) if recent_titles else "لا يوجد عناوين سابقة حالياً."
async def _gen_one(i: int) -> bool:
lang, vuln, diff = _pick_rotation_slot(int(time.time()) + i)
print(f" [{team_role}] Generating: {lang} / {vuln} / {diff}")
challenge = await ai_generate_challenge(lang, vuln, diff, blacklisted_titles=blacklisted_titles_str)
if challenge:
if await _insert_to_db(challenge):
print(f" [+] Inserted AI: {challenge['title']}")
return True
seeds = _build_seeds()
matching = [s for s in seeds if s["difficulty"] == diff]
if not matching:
matching = seeds
seed = random.choice(matching)
rand_tag = uuid.uuid4().hex[:6]
seed = dict(seed)
seed["title"] = f"{seed.get('title', 'تحدي ثغرة')} - رمز {rand_tag}"
seed["story"] = seed.get("story", "") + f" [معرف التدقيق: {rand_tag}]"
if await _insert_to_db(seed):
print(f" [+] Inserted seed fallback ({diff})")
return True
return False
results = []
for i in range(target):
ok = await _gen_one(i)
results.append(ok)
await asyncio.sleep(2.0)
return sum(1 for r in results if r)
async def refill_pool(team_role: str, count: int) -> int:
async with _get_pool_lock(team_role):
return await _refill_pool_inner(team_role, count)
# --------------------------------------------------------------------------- #
# 6. Pool Watcher
# --------------------------------------------------------------------------- #
async def start_pool_watcher(team_role: str):
if team_role not in ALLOWED_TEAMS:
return
if team_role in _WATCHER_STARTED:
print(f"[vuln-hunter:{team_role}] watcher already running — skipping duplicate start.")
return
_WATCHER_STARTED.add(team_role)
print(f"[vuln-hunter] Pool watcher started for '{team_role}' (target={POOL_TARGET}, threshold={POOL_THRESHOLD}, batch={POOL_BATCH})")
try:
while True:
try:
sleep_secs = 30
async with _get_pool_lock(team_role):
count = await get_pool_count(team_role)
if count <= POOL_THRESHOLD:
print(f"[vuln-hunter] Pool at {count}/{POOL_TARGET} (≤ {POOL_THRESHOLD}), refilling {POOL_BATCH}…")
added = await _refill_pool_inner(team_role, POOL_BATCH)
print(f"[vuln-hunter] Refill done: +{added} (now {count + added}/{POOL_TARGET})")
sleep_secs = 2
await asyncio.sleep(sleep_secs)
except Exception as e:
print(f"[vuln-hunter] Watcher error: {e}")
await asyncio.sleep(10)
finally:
_WATCHER_STARTED.discard(team_role)
# --------------------------------------------------------------------------- #
# 7. CLI
# --------------------------------------------------------------------------- #
def main():
parser = argparse.ArgumentParser(description="Vulnerability Hunter Challenge Generator")
parser.add_argument("--team", default="blue", choices=["blue"])
parser.add_argument("--ai", action="store_true", help="Use AI generation")
parser.add_argument("--count", type=int, default=3, help="Number to generate")
parser.add_argument("--seed-only", action="store_true", help="Insert seeds only")
args = parser.parse_args()
print(f"[vuln-hunter] Team: {args.team}, Count: {args.count}")
if args.seed_only:
seeds = _build_seeds()
random.shuffle(seeds)
inserted = 0
for seed in seeds[:args.count]:
import asyncio as _aio
if _aio.run(_insert_to_db(seed)):
inserted += 1
print(f" [+] Inserted seed: {seed['title']}")
print(f"[vuln-hunter] Seeded {inserted} challenges")
elif args.ai:
inserted = _aio.run(refill_pool(args.team, args.count))
print(f"[vuln-hunter] AI generated {inserted} challenges")
else:
_aio.run(start_pool_watcher(args.team))
if __name__ == "__main__":
main()