Cross-Site Scripting: DOM AI

Sending unvalidated data to a web browser can result in the browser executing malicious code.

Cross-site scripting (XSS) vulnerabilities occur when:

1. Data enters a web application through an untrusted source. In the case of DOM-based Artificial Intelegence (AI) XSS, data from an AI prompt response is written back into the page with client-side code. In the case of reflected XSS, the untrusted source is typically a web request, while in the case of persisted (also known as stored) XSS it is typically a database or other back-end data store.


2. The data is included in dynamic content that is sent to a web user without validation. In the case of DOM-based XSS, malicious content is executed as part of DOM (Document Object Model) creation, whenever the victim's browser parses the HTML page.

The malicious content sent to the web browser often takes the form of a JavaScript segment, but can also include HTML, Flash, or any other type of code that the browser executes. The variety of attacks based on XSS is almost limitless, but they commonly include transmitting private data such as cookies or other session information to the attacker, redirecting the victim to web content controlled by the attacker, or performing other malicious operations on the user's machine under the guise of the vulnerable site.

Example 1: The following JavaScript code segment retrieves a response from a large language model (LLM) and displays it to the user.

<SCRIPT>
var response = llm.invoke(user_prompt);
...
document.write(response);
</SCRIPT>

DOM-based XSS occurs when an attacker can modify a browser's DOM so that client-side code will be executed in an unintended, and potentially dangerous, manner.

The output of an LLM is unpredictable, enabling an attacker utilizing prompt engineering techniques to produce unexpected outputs and bypass developer instructions intended to prevent the LLM from producing dangerous responses.

As the example demonstrates, XSS vulnerabilities are caused by code that includes unvalidated responses from an LLM. There are three vectors by which an XSS attack can reach a victim:

- Data is read directly from the HTTP request and reflected back in the HTTP response. Reflected XSS exploits occur when an attacker causes a user to supply dangerous content to a vulnerable web application, which is then reflected back to the user and executed by the web browser. The most common mechanism for delivering malicious content is to include it as a parameter in a URL that is posted publicly or emailed directly to victims. URLs constructed in this manner constitute the core of many phishing schemes, whereby an attacker convinces victims to visit a URL that refers to a vulnerable site. After the site reflects the attacker's content back to the user, the content is executed and proceeds to transfer private information, such as cookies that might include session information, from the user's machine to the attacker or perform other nefarious activities.

- The application stores dangerous data in a database or other trusted data store. The dangerous data is subsequently read back into the application and included in dynamic content. Persistent XSS exploits occur when an attacker injects dangerous content into a data store that is later read and included in dynamic content. From an attacker's perspective, the optimal place to inject malicious content is in an area that is displayed to either many users or particularly interesting users. Interesting users typically have elevated privileges in the application or interact with sensitive data that is valuable to the attacker. If one of these users executes malicious content, the attacker can perform privileged operations on behalf of the user or gain access to sensitive data belonging to the user.

- A source outside the application produces potentially dangrous data, such as an LLM, stores dangerous data in a database, or retrieves data from another data store. This dangerous data is subsequently read back into the application as trusted data and included in dynamic content.