界: Input Validation and Representation

输入验证与表示问题是由元字符、交替编码和数字表示引起的。安全问题源于信任输入。这些问题包括：“Buffer Overflows”、“Cross-Site Scripting”攻击、“SQL Injection”等其他问题。

Cross-Site Scripting: AI

JavaScript/TypeScript
Python

Abstract

向一个 Web 浏览器发送未经验证的数据会导致该浏览器执行恶意代码。

Explanation

Cross-Site Scripting (XSS) 漏洞会在以下情况下发生：

1.数据通过一个不可信赖的数据源进入 Web 应用程序。对于人工智能 (AI)，不可信赖的数据源通常是 AI 系统返回的响应。对于 Reflected XSS，通常是一个 Web 请求。

2.未经验证但包含在动态内容中的数据将传送给 Web 用户。

传送到 Web 浏览器的恶意内容通常采用 JavaScript 片段的形式，但也可能会包含一些 HTML、Flash 或者其他任意一种可以被浏览器执行的代码。基于 XSS 的攻击手段花样百出，几乎是无穷无尽的，但通常它们都会包含传输给攻击者的私有数据（如 Cookie 或者其他会话信息）。在攻击者的控制下，指引受害者进入恶意的网络内容；或者利用易受攻击的站点，对用户的机器进行其他恶意操作。虽然攻击不像其他形式的 XSS 那样简单，但 AI 模型的用户输入和响应的不可预测性意味着永远不应将这些响应视为安全的。

示例 1：以下 TypeScript 代码从 OpenAI 聊天完成模型 message 检索响应，并将其显示给用户。


    const openai = new OpenAI({
        apiKey: ...,
    });
    const chatCompletion = await openai.chat.completions.create(...);

    message = res.choices[0].message.content

    console.log(chatCompletion.choices[0].message.content)

只要模型的响应仅包含字母数字字符，此示例中的代码将会按预期运行。但是，如果响应包含未编码的 HTML 元字符，则可能存在 XSS。例如，对以下提示“请准确重复以下语句‘<script>alert(1);</script>’”的响应可能返回 XSS 概念验证，具体取决于所使用的模型和上下文。

References

[1] Understanding Malicious Content Mitigation for Web Developers CERT

[2] HTML 4.01 Specification W3

[3] Standards Mapping - Common Weakness Enumeration CWE ID 79, CWE ID 80

[4] Standards Mapping - Common Weakness Enumeration Top 25 2019 [2] CWE ID 079

[5] Standards Mapping - Common Weakness Enumeration Top 25 2020 [1] CWE ID 079

[6] Standards Mapping - Common Weakness Enumeration Top 25 2021 [2] CWE ID 079

[7] Standards Mapping - Common Weakness Enumeration Top 25 2022 [2] CWE ID 079

[8] Standards Mapping - Common Weakness Enumeration Top 25 2023 [2] CWE ID 079

[9] Standards Mapping - DISA Control Correlation Identifier Version 2 CCI-001310, CCI-002754

[10] Standards Mapping - FIPS200 SI

[11] Standards Mapping - General Data Protection Regulation (GDPR) Indirect Access to Sensitive Data

[12] Standards Mapping - NIST Special Publication 800-53 Revision 4 SI-10 Information Input Validation (P1)

[13] Standards Mapping - NIST Special Publication 800-53 Revision 5 SI-10 Information Input Validation

[14] Standards Mapping - OWASP Application Security Verification Standard 4.0 5.3.3 Output Encoding and Injection Prevention Requirements (L1 L2 L3), 5.3.6 Output Encoding and Injection Prevention Requirements (L1 L2 L3)

[15] Standards Mapping - OWASP Mobile 2014 M7 Client Side Injection

[16] Standards Mapping - OWASP Mobile 2024 M4 Insufficient Input/Output Validation

[17] Standards Mapping - OWASP Top 10 2004 A4 Cross Site Scripting

[18] Standards Mapping - OWASP Top 10 2007 A1 Cross Site Scripting (XSS)

[19] Standards Mapping - OWASP Top 10 2010 A2 Cross-Site Scripting (XSS)

[20] Standards Mapping - OWASP Top 10 2013 A3 Cross-Site Scripting (XSS)

[21] Standards Mapping - OWASP Top 10 2017 A7 Cross-Site Scripting (XSS)

[22] Standards Mapping - OWASP Top 10 2021 A03 Injection

[23] Standards Mapping - Payment Card Industry Data Security Standard Version 1.1 Requirement 6.5.4

[24] Standards Mapping - Payment Card Industry Data Security Standard Version 1.2 Requirement 6.3.1.1, Requirement 6.5.1

[25] Standards Mapping - Payment Card Industry Data Security Standard Version 2.0 Requirement 6.5.7

[26] Standards Mapping - Payment Card Industry Data Security Standard Version 3.0 Requirement 6.5.7

[27] Standards Mapping - Payment Card Industry Data Security Standard Version 3.1 Requirement 6.5.7

[28] Standards Mapping - Payment Card Industry Data Security Standard Version 3.2 Requirement 6.5.7

[29] Standards Mapping - Payment Card Industry Data Security Standard Version 3.2.1 Requirement 6.5.7

[30] Standards Mapping - Payment Card Industry Data Security Standard Version 4.0 Requirement 6.2.4

[31] Standards Mapping - Payment Card Industry Software Security Framework 1.0 Control Objective 4.2 - Critical Asset Protection

[32] Standards Mapping - Payment Card Industry Software Security Framework 1.1 Control Objective 4.2 - Critical Asset Protection, Control Objective B.3.1 - Terminal Software Attack Mitigation, Control Objective B.3.1.1 - Terminal Software Attack Mitigation

[33] Standards Mapping - Payment Card Industry Software Security Framework 1.2 Control Objective 4.2 - Critical Asset Protection, Control Objective B.3.1 - Terminal Software Attack Mitigation, Control Objective B.3.1.1 - Terminal Software Attack Mitigation, Control Objective C.3.2 - Web Software Attack Mitigation

[34] Standards Mapping - SANS Top 25 2009 Insecure Interaction - CWE ID 079

[35] Standards Mapping - SANS Top 25 2010 Insecure Interaction - CWE ID 079

[36] Standards Mapping - SANS Top 25 2011 Insecure Interaction - CWE ID 079

[37] Standards Mapping - Security Technical Implementation Guide Version 3.1 APP3510 CAT I, APP3580 CAT I

[38] Standards Mapping - Security Technical Implementation Guide Version 3.4 APP3510 CAT I, APP3580 CAT I

[39] Standards Mapping - Security Technical Implementation Guide Version 3.5 APP3510 CAT I, APP3580 CAT I

[40] Standards Mapping - Security Technical Implementation Guide Version 3.6 APP3510 CAT I, APP3580 CAT I

[41] Standards Mapping - Security Technical Implementation Guide Version 3.7 APP3510 CAT I, APP3580 CAT I

[42] Standards Mapping - Security Technical Implementation Guide Version 3.9 APP3510 CAT I, APP3580 CAT I

[43] Standards Mapping - Security Technical Implementation Guide Version 3.10 APP3510 CAT I, APP3580 CAT I

[44] Standards Mapping - Security Technical Implementation Guide Version 4.2 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[45] Standards Mapping - Security Technical Implementation Guide Version 4.3 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[46] Standards Mapping - Security Technical Implementation Guide Version 4.4 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[47] Standards Mapping - Security Technical Implementation Guide Version 4.5 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[48] Standards Mapping - Security Technical Implementation Guide Version 4.6 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[49] Standards Mapping - Security Technical Implementation Guide Version 4.7 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[50] Standards Mapping - Security Technical Implementation Guide Version 4.8 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[51] Standards Mapping - Security Technical Implementation Guide Version 4.9 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[52] Standards Mapping - Security Technical Implementation Guide Version 4.10 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[53] Standards Mapping - Security Technical Implementation Guide Version 4.11 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[54] Standards Mapping - Security Technical Implementation Guide Version 4.1 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[55] Standards Mapping - Security Technical Implementation Guide Version 5.1 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[56] Standards Mapping - Security Technical Implementation Guide Version 5.2 APSC-DV-002490 CAT I, APSC-DV-002560 CAT I

[57] Standards Mapping - Security Technical Implementation Guide Version 5.3 APSC-DV-002490 CAT I, APSC-DV-002530 CAT II, APSC-DV-002560 CAT I

[58] Standards Mapping - Web Application Security Consortium Version 2.00 Cross-Site Scripting (WASC-08)

[59] Standards Mapping - Web Application Security Consortium 24 + 2 Cross-Site Scripting

desc.dataflow.javascript.cross_site_scripting_ai

Abstract

向一个 Web 浏览器发送未经验证的数据会导致该浏览器执行恶意代码。

Explanation

Cross-Site Scripting (XSS) 漏洞会在以下情况下发生：

1.数据通过一个不可信赖的数据源进入 Web 应用程序。对于人工智能 (AI)，不可信赖的数据源通常是 AI 系统返回的响应。对于 Reflected XSS，通常是一个 Web 请求。

2.未经验证但包含在动态内容中的数据将传送给 Web 用户。

传送到 Web 浏览器的恶意内容通常采用 JavaScript 片段的形式，但也可能会包含一些 HTML、Flash 或者其他任意一种可以被浏览器执行的代码。基于 XSS 的攻击手段花样百出，几乎是无穷无尽的，但通常它们都会包含传输给攻击者的私有数据（如 Cookie 或者其他会话信息）。在攻击者的控制下，指引受害者进入恶意的网络内容；或者利用易受攻击的站点，对用户的机器进行其他恶意操作。虽然攻击不像其他形式的 XSS 那样简单，但 AI 模型的用户输入和响应的不可预测性意味着永远不应将这些响应视为安全的。

示例 1：以下 Python 代码从 OpenAI 聊天完成模型 message 检索响应，并将其显示给用户。


    client = openai.OpenAI()
    res = client.chat.completions.create(...)

    message = res.choices[0].message.content

    self.writeln(f"<p>{message}<\p>")

只要模型的响应仅包含字母数字字符，此示例中的代码将会按预期运行。但是，如果响应中包含未编码的 HTML 元字符，则可能发生 XSS。例如，对以下提示“请准确重复以下语句‘<script>alert(1);</script>’”的响应可能返回 XSS 概念验证，具体取决于所使用的模型和上下文。