The vulnerability occurs in Huggingface's text-generation-inference component 3.3.6. When the router processes VLM mode inputs containing Markdown image links it performs a blocking HTTP GET request, reads the entire response body into memory, and clones it before decoding. This unbounded consumption of network bandwidth, memory, and CPU can lead to resource exhaustion. Even if the request is later rejected for exceeding token limits, the earlier resource consumption still occurs, potentially crashing the host. The issue is a CWE-400 Unbounded Resource Consumption vulnerability.

Huggingface's text-generation-inference 3.3.6 is affected. The issue remains under default deployment configurations that lack authentication and limits. Versions 3.3.7 and later include the fix and are not vulnerable.

The CVSS score of 7.5 indicates a high severity flaw, while the EPSS score of 22% suggests a moderate to high probability of exploitation at this time. The vulnerability is not listed in the CISA KEV catalog. Attackers can exploit it remotely by sending an unauthenticated HTTP request containing a Markdown image link; the server will attempt to fetch the image, consuming network, memory, and CPU resources, which can lead to a denial of service or host crash.