漏洞简介

vLLM 是大型语言模型(LLM)的推理和服务引擎。在 0.14.1 版本之前,vLLM 项目多模态功能集中的“MediaConnector”类别存在SSRF漏洞。load_from_url和load_from_url_async方法通过用户提供的URL获取和处理媒体,在限制目标主机时使用不同的Python解析库。这两个解析库对反斜线有不同的解释,这使得可以绕过主机名的限制。这使得攻击者能够胁迫vLLM服务器对内部网络资源发出任意请求。该漏洞在容器化环境中尤为关键,如“LLM-d”,其中被攻破的vLLM胶囊可能被用来扫描内部网络、与其他胶囊交互,甚至可能导致拒绝服务或访问敏感数据。例如,攻击者可能会让vLLM播客向内部的“LLM-d”管理端点发送恶意请求,从而通过错误报告如KV缓存状态等指标导致系统不稳定。0.14.1版本包含该问题的补丁。

漏洞影响

评分:7.1

影响:<0.14.1

漏洞分析

前端请求

POST /v1/chat/completions HTTP/1.1
Authorization: Bearer <valid_token>
Content-Type: application/json

{
    "model": "llava",
    "messages": [
        {
            "role": "user", 
            "content": [
                {
                    "type": "text",
                    "text": "请分析这张图片的内容"
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "http://allowed.com\\@192.168.1.100:8080/internal-api/admin"
                    }
                }
            ]
        }
    ],
    "max_tokens": 100
}

后端调用链

这里开始响应

handler = chat(raw_request)

获取处理器

走到chat方法

从当前应用的全局状态中取出名为 openai_serving_chat 的对象,并返回它。该属性的值是一个 OpenAIServingChat 类的实例

generator = await handler.create_chat_completion(request, raw_request)

调用 handler 的 异步方法 create_chat_completion

实际上就是OpenAIServingChat.create_chat_completion

跳到create_chat_completion函数处

conversation, engine_prompts = await self._preprocess_chat(
                    request,
                    tokenizer,
                    request.messages,
                    chat_template=request.chat_template or self.chat_template,
                    chat_template_content_format=self.chat_template_content_format,
                    add_generation_prompt=request.add_generation_prompt,
                    continue_final_message=request.continue_final_message,
                    tool_dicts=tool_dicts,
                    documents=request.documents,
                    chat_template_kwargs=chat_template_kwargs,
                    default_chat_template_kwargs=self.default_chat_template_kwargs,
                    tool_parser=tool_parser,
                    add_special_tokens=request.add_special_tokens,
                )

开始预处理消息,其中request.messages包含了前端传入的恶意url

跳到_preprocess_chat看看如何进行消息预处理的

conversation, mm_data_future, mm_uuids = parse_chat_messages_futures(
            messages,
            model_config,
            content_format=resolved_content_format,
        )
        mm_data = await mm_data_future

解析消息,创建异步媒体获取任务,parse_chat_messages_futures方法接收了含恶意URL的messages

mm_data等待媒体数据

跳到parse_chat_messages_futures方法看看

mm_tracker = AsyncMultiModalItemTracker(model_config)
    for msg in messages:
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
            content_format,
            interleave_strings=(
                content_format == "string"
                and model_config.multimodal_config is not None
                and model_config.multimodal_config.interleave_mm_strings
            ),
        )

         return conversation, mm_tracker.all_mm_data(), mm_tracker.all_mm_uuids()

先创建mm_tracker跟踪器

使用for循环每次循环传入msg(恶意url也被传入),该参数表示一条消息

然后mm_tracker参数传递跟踪器

跳到_parse_chat_message_content方法

elif isinstance(content, str):
        content = [ChatCompletionContentPartTextParam(type="text", text=content)]
    result = _parse_chat_message_content_parts(
        role,
        content,  # type: ignore
        mm_tracker,
        wrap_dicts=(content_format == "openai"),
        interleave_strings=interleave_strings,
    )

解析内容部分

content为内容数组:[text, image_url](将恶意url传递下去),mm_tracker参数继续传递跟踪器

再跳到_parse_chat_message_content_parts方法

for part in parts:
        parse_res = _parse_chat_message_content_part(
            part,
            mm_parser,
            wrap_dicts=wrap_dicts,
            interleave_strings=interleave_strings,
        )

使用for循环循环调用 _parse_chat_message_content_part方法,每次使用传参part(将恶意url传递下去),该参数表示单个内容部分

mm_parser继续传递解析器

跳到_parse_chat_message_content_part方法

elif part_type in ("image_url", "input_image"):
        str_content = cast(str, content)
        mm_parser.parse_image(str_content, uuid)
        modality = "image"

识别 image_url 类型,提取 URL,调用图片解析器

提取出恶意URL字符串传递给解析器mm_parser的parse_image

跳到parse_image方法

调用self._connector.fetch_image_async(image_url)获取图片数据,创建占位符引用

跳到fetch_image_async方法

return await self.load_from_url_async(
                image_url,
                image_io,
                fetch_timeout=envs.VLLM_IMAGE_FETCH_TIMEOUT,
            )

调用load_from_url_async加载URL也就是我们的恶意url


跳到load_from_url_async方法

url_spec = urlparse(url)
        loop = asyncio.get_running_loop()

        if url_spec.scheme.startswith("http"):
            self._assert_url_in_allowed_media_domains(url_spec)

            connection = self.connection
            data = await connection.async_get_bytes(
                url,
                timeout=fetch_timeout,
                allow_redirects=envs.VLLM_MEDIA_URL_ALLOW_REDIRECTS,
            )
            future = loop.run_in_executor(global_thread_pool, media_io.load_bytes, data)
            return await future

使用 urlparse,这是python标准库,遵循:RFC 3986 规范:将反斜杠 `\` 视为普通字符,对某些特殊字符处理严格

进行解析我们的恶意URL

然后经过白名单

ister(global_thread_pool.shutdown)

_M = TypeVar("_M")

MEDIA_CONNECTOR_REGISTRY = ExtensionManager()


@MEDIA_CONNECTOR_REGISTRY.register("http")
class MediaConnector:
    def __init__(
        self,
        media_io_kwargs: dict[str, dict[str, Any]] | None = None,
        connection: HTTPConnection = global_http_connection,
        *,
        allowed_local_media_path: str = "",
        allowed_media_domains: list[str] | None = None,
    ) -> None:
        """
        Args:
            media_io_kwargs: Additional args passed to process media
                             inputs, keyed by modalities. For example,
                             to set num_frames for video, set
                             `--media-io-kwargs '{"video":{"num_frames":40}}'`
            connection: HTTP connection client to download media contents.
            allowed_local_media_path: A local directory to load media files from.
            allowed_media_domains: If set, only media URLs that belong to this
                                   domain can be used for multi-modal inputs.
        """
        super().__init__()

        self.media_io_kwargs: dict[str, dict[str, Any]] = (
            media_io_kwargs if media_io_kwargs else {}
        )
        self.connection = connection

        if allowed_local_media_path:
            allowed_local_media_path_ = Path(allowed_local_media_path)

            if not allowed_local_media_path_.exists():
                raise ValueError(
                    "Invalid `--allowed-local-media-path`: The path "
                    f"{allowed_local_media_path_} does not exist."
                )
            if not allowed_local_media_path_.is_dir():
                raise ValueError(
                    "Invalid `--allowed-local-media-path`: The path "
                    f"{allowed_local_media_path_} must be a directory."
                )
        else:
            allowed_local_media_path_ = None

        self.allowed_local_media_path = allowed_local_media_path_
        if allowed_media_domains is None:
            allowed_media_domains = []
        self.allowed_media_domains = allowed_media_domains

urlparse解析恶意payload异常,导致

url_spec.hostname = None 或 ""

白名单检查时

if False and True → False

从而跳过检查

最后调用HTTP连接器发起实际请求

跳到async_get_bytes方法

async with await self.get_async_response(
            url, timeout=timeout, allow_redirects=allow_redirects
        ) as r:
            r.raise_for_status()

            return await r.read()

调用get_async_response方法

检查HTTP状态码并读取返回响应体

跳到get_async_response方法

self._validate_http_url(url)

进行了URL验证

跳到该验证处

parsed_url = urlparse(url)

再次使用urlparse

在这里,只检查了scheme,其他的像检查主机名,白名单验证,内部地址检查什么的都没有

通过验证处之后

client = await self.get_async_client()
        extra_headers = extra_headers or {}

        return client.get(
            url,
            headers=self._headers(**extra_headers),
            timeout=timeout,
            allow_redirects=allow_redirects,
        )

当调用 client.get() 时:

client 是 aiohttp.ClientSession 实例

get() 是异步方法,返回协程对象,然后对我们的恶意URL发起网络请求
不过该
URL解析使用的方法是parse_url,按照WHATWG生活标准实现:将 @ 符号识别为认证信息分隔符,即使前面有反斜杠。

跳到self.get_async_client看看

可以看到client 是 aiohttp.ClientSession 实例

漏洞修复

统一使用 parse_url解析url

更新 _assert_url_in_allowed_media_domains 方法签名: 验证URL主机名是否在允许的域名列表中

参考

https://github.com/vllm-project/vllm/security/advisories/GHSA-qh4c-xf7m-gxfc

https://www.cve.org/CVERecord?id=CVE-2026-24779

https://github.com/vllm-project/vllm/tree/v0.14.0/vllm

https://github.com/vllm-project/vllm/commit/f46d576c54fb8aeec5fc70560e850bed38ef17d7