CVE-2025-46560: vLLM: DoS via quadratic multimodal tokenizer input

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Versions starting from 0.8.0 and prior to 0.8.5 are affected by a critical performance vulnerability in the input preprocessing logic of the multimodal tokenizer. The code dynamically replaces placeholder tokens (e.g., <|audio_|>, <|image_|>) with repeated tokens based on precomputed lengths. Due to ​​inefficient list concatenation operations​​, the algorithm exhibits ​​quadratic time complexity (O(n²))​​, allowing malicious actors to trigger resource exhaustion via specially crafted inputs. This issue has been patched in version 0.8.5.

An adversary identifies a public-facing LLM API powered by vLLM 0.8.x (discoverable via model metadata endpoints, HTTP headers, or open-source deployment docs). They craft a multimodal request containing an abnormally large sequence of image/audio placeholder tokens (e.g., hundreds of <|image_1|> tokens) and submit it to the inference endpoint. The tokenizer's quadratic list concatenation causes processing time to explode — what should take milliseconds takes seconds or minutes — exhausting CPU and memory on the inference worker. The attacker sends a modest volume of such requests concurrently, causing the serving process to stall or OOM-crash. In a Kubernetes deployment this may trigger cascading pod restarts; in a bare-metal deployment it takes the inference service offline. No credentials, no prior access, no AI/ML expertise required beyond knowing the placeholder token format, which is documented in the vLLM public codebase.