An unauthenticated remote denial of service vulnerability exists in llama.cpp when the --reranking flag is enabled, posing a risk to service availability.

This is a denial of service vulnerability triggered by an unauthenticated attacker sending a crafted POST request to the /rerank endpoint containing a negative top_n parameter. This input causes an unhandled memory allocation error (std::bad_alloc) and a subsequent HTTP 500 response, potentially crashing the application process.

The exploitation of this vulnerability results in service downtime, which can disrupt business operations dependent on the llama.cpp inference engine. With a CVSS score of 7.5, the vulnerability is classified as high severity because it is network-accessible and requires no authentication, allowing any remote user to impact the availability of the service.

Immediate Action: Ensure the --reranking flag is disabled if not strictly required, and monitor the upstream repository for a commit that addresses the memory allocation error.

Proactive Monitoring: Review application and web server access logs for anomalous POST requests to the /rerank endpoint, specifically looking for negative integers passed to the top_n parameter.

Compensating Controls: Deploy a Web Application Firewall (WAF) or an API gateway to sanitize incoming requests and block POST traffic to /rerank that contains negative values in the top_n field.

Public Exploit Available: Unknown.

Given the high CVSS score and the ease of exploitation over the network, organizations using llama.cpp with the reranking feature enabled must treat this as a significant availability risk. Organizations should prioritize restricting access to the /rerank endpoint or disabling the feature until a formal security patch is released by the project maintainers.