CVE-2024-1522

A Cross-Site Request Forgery (CSRF) vulnerability in the parisneo/lollms-webui project allows remote attackers to execute arbitrary code on a victim's system. The vulnerability stems from the /execute_code API endpoint, which does not properly validate requests, enabling an attacker to craft a malicious webpage that, when visited by a victim, submits a form to the victim's local lollms-webui instance to execute arbitrary OS commands. This issue allows attackers to take full control of the victim's system without requiring direct network access to the vulnerable application.

A Cross-Site Request Forgery (CSRF) vulnerability in the parisneo/lollms-webui project allows remote attackers to execute arbitrary code on a victim's system. The vulnerability stems from the /execute_code API endpoint, which does not properly validate requests, enabling an attacker to craft a malicious webpage that, when visited by a victim, submits a form to the victim's local lollms-webui instance to execute arbitrary OS commands. This issue allows attackers to take full control of the victim's system without requiring direct network access to the vulnerable application.

I have activated the CORS because I had a development ui that uses another port number then I forgot to remove it.

So what I just did is :

• First removed the cors configuration that allows everyone to access it : before:

    sio = socketio.AsyncServer(async_mode="asgi", cors_allowed_origins="*", ping_timeout=1200, ping_interval=30)  # Enable CORS for every one

after:

    cert_file_path = lollms_paths.personal_certificates/"cert.pem"
    key_file_path = lollms_paths.personal_certificates/"key.pem"
    if os.path.exists(cert_file_path) and os.path.exists(key_file_path):
        is_https = True
    else:
        is_https = False        

    # Create a Socket.IO server
    sio = socketio.AsyncServer(async_mode="asgi", cors_allowed_origins=config.allowed_origins+[f"https://localhost:{config['port']}" if is_https else f"http://localhost:{config['port']}"], ping_timeout=1200, ping_interval=30)  # Enable CORS for selected origins

• Second, I have updated lollms to have two modes (a headless mode and a ui mode). And updated the /execute_code to block if the server is headless or is exposed

@router.post("/execute_code")
async def execute_code(request: Request):
    """
    Executes Python code and returns the output.

    :param request: The HTTP request object.
    :return: A JSON response with the status of the operation.
    """
    if lollmsElfServer.config.headless_server_mode:
        return {"status":False,"error":"Code execution is blocked when in headless mode for obvious security reasons!"}

    if lollmsElfServer.config.host=="0.0.0.0":
        return {"status":False,"error":"Code execution is blocked when the server is exposed outside for very obvipous reasons!"}

    try:
        data = (await request.json())
        code = data["code"]
        discussion_id = int(data.get("discussion_id","unknown_discussion"))
        message_id = int(data.get("message_id","unknown_message"))
        language = data.get("language","python")
        


        if language=="python":
            ASCIIColors.info("Executing python code:")
            ASCIIColors.yellow(code)
            return execute_python(code, discussion_id, message_id)
        if language=="javascript":
            ASCIIColors.info("Executing javascript code:")
            ASCIIColors.yellow(code)
            return execute_javascript(code, discussion_id, message_id)
        if language in ["html","html5","svg"]:
            ASCIIColors.info("Executing javascript code:")
            ASCIIColors.yellow(code)
            return execute_html(code, discussion_id, message_id)
        
        elif language=="latex":
            ASCIIColors.info("Executing latex code:")
            ASCIIColors.yellow(code)
            return execute_latex(code, discussion_id, message_id)
        elif language in ["bash","shell","cmd","powershell"]:
            ASCIIColors.info("Executing shell code:")
            ASCIIColors.yellow(code)
            return execute_bash(code, discussion_id, message_id)
        elif language in ["mermaid"]:
            ASCIIColors.info("Executing mermaid code:")
            ASCIIColors.yellow(code)
            return execute_mermaid(code, discussion_id, message_id)
        elif language in ["graphviz","dot"]:
            ASCIIColors.info("Executing graphviz code:")
            ASCIIColors.yellow(code)
            return execute_graphviz(code, discussion_id, message_id)
        return {"status": False, "error": "Unsupported language", "execution_time": 0}
    except Exception as ex:
        trace_exception(ex)
        lollmsElfServer.error(ex)
        return {"status":False,"error":str(ex)}

I also added an optional https mode and looking forward to add a full authentication with cookies and a personal session etc.

All updates will be in V 9.1

Again, thanks alot for your work. I will make it harder next time, but if you find more bugs, just be my guest :)

A Cross-Site Request Forgery (CSRF) vulnerability in the parisneo/lollms-webui project allows remote attackers to execute arbitrary code on a victim's system. The vulnerability stems from the /execute_code API endpoint, which does not properly validate requests, enabling an attacker to craft a malicious webpage that, when visited by a victim, submits a form to the victim's local lollms-webui instance to execute arbitrary OS commands. This issue allows attackers to take full control of the victim's system without requiring direct network access to the vulnerable application.

A Cross-Site Request Forgery (CSRF) vulnerability in the parisneo/lollms-webui project allows remote attackers to execute arbitrary code on a victim's system. The vulnerability stems from the /execute_code API endpoint, which does not properly validate requests, enabling an attacker to craft a malicious webpage that, when visited by a victim, submits a form to the victim's local lollms-webui instance to execute arbitrary OS commands. This issue allows attackers to take full control of the victim's system without requiring direct network access to the vulnerable application.