CVE-2024-8768 is a reachable assertion vulnerability identified in the vLLM library, specifically triggered when a completions API request is made with an empty prompt. The flaw causes the vLLM API server to crash, leading to a denial of service (DoS) condition. The vulnerability is classified with a CVSS 3.1 base score of 7.5, indicating high severity. The vector metrics indicate that the attack can be performed remotely (Network), with low attack complexity, no privileges required, and no user interaction needed. The scope remains unchanged, and the impact is limited to availability, with no confidentiality or integrity loss. The vulnerability affects version 0 of the vLLM library, which is used in AI and machine learning applications to provide language model completions. The root cause is an assertion failure triggered by an empty prompt input, which is not properly handled by the server, causing it to crash. There are no patches or known exploits publicly available at this time, but the vulnerability is published and should be addressed promptly. The flaw can be exploited by sending crafted API requests, making it a straightforward vector for denial of service attacks against services relying on vLLM.

The primary impact of CVE-2024-8768 is denial of service, which can disrupt availability of AI-powered services and applications that depend on the vLLM library for language model completions. Organizations using vLLM in production environments may experience service outages or degraded performance if attackers exploit this vulnerability by sending empty prompt requests. This can affect customer-facing applications, internal tools, or any automated systems relying on vLLM, potentially leading to operational downtime and loss of user trust. While the vulnerability does not compromise data confidentiality or integrity, the availability impact can have cascading effects, including delayed business processes, reduced productivity, and potential financial losses. The ease of exploitation and lack of required privileges increase the risk of widespread abuse, especially in environments exposed to untrusted networks or the internet. The absence of known exploits in the wild currently limits immediate risk but does not preclude future exploitation.

To mitigate CVE-2024-8768, organizations should first check for and apply any official patches or updates released by the vLLM maintainers once available. In the absence of patches, implement input validation at the API gateway or application layer to reject completions requests with empty prompts before they reach the vLLM server. Employ rate limiting and anomaly detection to identify and block suspicious traffic patterns that may indicate exploitation attempts. Consider deploying web application firewalls (WAFs) with custom rules to filter out malformed or empty prompt requests targeting the completions API. Additionally, isolate the vLLM service within a segmented network zone to limit exposure and reduce the blast radius of potential attacks. Monitoring and logging API request patterns can help detect exploitation attempts early. Finally, prepare incident response plans to quickly restore service availability in case of a successful denial of service attack.