{"id":"CVE-2026-115916","kind":"vulnerability","aliases":["CVE-2026-115916"],"related":[],"title":"vLLM: an out-of-range prompt token ID crashes the serving engine","description":"vLLM is an inference and serving engine for large language models. Prior to 0.32.0, vllm/tokenizers/hf.py computes max_token_id with an off-by-one upper bound, allowing a pre-tokenized prompt ID equal to the vocabulary size to pass validation. For a tensor-parallel size of one and an exact multiple-of-64 vocabulary, VocabParallelEmbedding.forward does not mask that value before F.embedding uses it, causing a CUDA device-side assertion and terminating the engine. This issue is fixed in version 0.32.0.","summary":"A pre-tokenized prompt containing a token ID equal to the vocabulary size passes validation in vLLM before 0.32.0. On a single-GPU setup with a vocabulary that is an exact multiple of 64, the embedding layer hits a CUDA device-side assertion and the engine process terminates, which is a denial of service.","severity":"medium","cvss":6.5,"epss":null,"exploited":false,"exploited_at":null,"products":[],"ecosystems":[],"tags":["ai-tooling","gpu-inference","python","self-hosted","vllm"],"not_affected_if":["managed-hosting","text-prompts-only"],"how_to_check":"- Run `python -c \"import vllm; print(vllm.__version__)\"` and check whether the result is below `0.32.0`\n- Check whether the deployment runs with tensor-parallel size `1` and serves a model whose vocabulary size is an exact multiple of 64\n- Check whether clients can send pre-tokenized prompt IDs directly to the server","fix":"- Upgrade vLLM to `0.32.0` or later\n- Until upgraded, limit who can submit requests to the serving endpoint, since the crash needs a crafted token ID","status":"open","published_at":"2026-10-26T18:17:39+00:00","first_seen_at":"2026-10-26T18:31:43+00:00","triaged_at":"2026-10-26T18:34:01+00:00","updated_at":"2026-10-26T18:34:01+00:00","category":"ai-tooling","confidence":0.7,"triage":{"at":"2026-10-26T18:34:01Z","model":"claude-haiku-5-5","output":{"fix":"- Upgrade vLLM to `0.32.0` or later\n- Until upgraded, limit who can submit requests to the serving endpoint, since the crash needs a crafted token ID","reach":"network","title":"vLLM: an out-of-range prompt token ID crashes the serving engine","impact":"low","summary":"A pre-tokenized prompt containing a token ID equal to the vocabulary size passes validation in vLLM before 0.32.0. On a single-GPU setup with a vocabulary that is an exact multiple of 64, the embedding layer hits a CUDA device-side assertion and the engine process terminates, which is a denial of service.","category":"ai-tooling","nicknames":[],"who_cares":["vllm","self-hosted","gpu-inference","python"],"confidence":0.7,"needs_auth":"unknown","worst_case":"Inference engine process crashes, causing denial of service until restarted","analyst_note":"","how_to_check":"- Run `python -c \"import vllm; print(vllm.__version__)\"` and check whether the result is below `0.32.0`\n- Check whether the deployment runs with tensor-parallel size `1` and serves a model whose vocabulary size is an exact multiple of 64\n- Check whether clients can send pre-tokenized prompt IDs directly to the server","needs_analyst":false,"not_affected_if":["managed-hosting","text-prompts-only"],"who_is_affected":"Teams self-hosting vLLM for inference on a single GPU (tensor-parallel size one) with a model whose vocabulary size is an exact multiple of 64. The flaw needs a request that supplies pre-tokenized prompt IDs, so an endpoint that only accepts text prompts is less exposed. Teams using hosted model APIs or multi-GPU tensor-parallel setups are less likely to be hit.","user_interaction":"no","needs_special_setup":"yes","severity_if_unknown":"medium"},"cost_usd":0.000671,"input_tokens":3388,"output_tokens":665,"severity_before":"medium"},"analyst_status":null,"analyst_note":null,"search":"'-115916':16A,19A '-2026':15A,18A '0.32.0':50B,101C,173C '64':66B,144C 'ai':21B 'ai-tool':20B 'allow':115C 'assert':76B,162C 'bound':114C 'caus':156C 'comput':103C 'contain':36B 'crash':10A 'cuda':72B,158C 'cve':14A,17A 'denial':85B 'devic':74B,160C 'device-sid':73B,159C 'embed':68B 'engin':13A,79B,94C,166C 'equal':40B,122C 'exact':63B,140C 'f.embedding':153C 'fix':170C 'gpu':24B,55B 'gpu-infer':23B 'hit':70B 'host':29B 'id':9A,39B,106C,121C 'infer':25B,91C 'issu':168C 'languag':97C 'larg':96C 'layer':69B 'mask':149C 'max':104C 'model':98C 'multipl':64B,142C 'multiple-of':141C 'off-by-on':109C 'one':112C,137C 'out-of-rang':3A 'parallel':134C 'pass':45B,128C 'pre':33B,118C 'pre-token':32B,117C 'prior':99C 'process':80B 'prompt':7A,35B,120C 'python':26B 'rang':6A 'self':28B 'self-host':27B 'serv':12A,93C 'servic':87B 'setup':56B 'side':75B,161C 'singl':54B 'single-gpu':53B 'size':44B,126C,135C 'tensor':133C 'tensor-parallel':132C 'termin':81B,164C 'token':8A,34B,38B,105C,119C 'tool':22B 'upper':113C 'use':154C 'valid':46B,129C 'valu':151C 'version':172C 'vllm':1A,30B,48B,88C 'vllm/tokenizers/hf.py':102C 'vocabparallelembedding.forward':146C 'vocabulari':43B,59B,125C,145C","nicknames":[],"mitigations":[],"who_affected":"Teams self-hosting vLLM for inference on a single GPU (tensor-parallel size one) with a model whose vocabulary size is an exact multiple of 64. The flaw needs a request that supplies pre-tokenized prompt IDs, so an endpoint that only accepts text prompts is less exposed. Teams using hosted model APIs or multi-GPU tensor-parallel setups are less likely to be hit.","tier":"B","verified":{},"analyst_revisit_at":null,"first_public_at":"2026-10-26T18:17:39+00:00","kev":null,"cvss_vector":"CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H","cvss_by":"security-advisories@github.com","poc_count":0,"sources":[{"source":"nvd","external_id":"CVE-2026-115916","url":"https://nvd.nist.gov/vuln/detail/CVE-2026-115916","severity":null,"published_at":"2026-10-26T18:17:39+00:00","modified_at":"2026-10-26T18:17:39+00:00","first_seen_at":"2026-10-26T18:31:43+00:00","fetched_at":"2026-10-26T18:31:43+00:00","note":null}],"documents":[],"notice":"Free to read at a human pace. An API with keys and an MCP server are in preparation; early access: hello@infrastack.ai"}