{"id":"CVE-2026-116502","kind":"vulnerability","aliases":["CVE-2026-116502"],"related":[],"title":"ExLlamaV2: crafted g_idx input causes out-of-bounds write locally","description":"A flaw has been found in turboderp-org ExLlamaV2 up to 0.3.2. The impacted element is the function QMatrix::make_sequential of the file exllamav2/exllamav2_ext/cuda/q_matrix.cu of the component GPTQ Index Handler. Executing a manipulation of the argument g_idx can lead to out-of-bounds write. The attack needs to be launched locally. The project was informed of the problem early through an issue report but has not responded yet.","summary":"ExLlamaV2 up to 0.3.2 has an out-of-bounds write in the GPTQ index handling code (QMatrix::make_sequential in the CUDA extension) when the g_idx argument is manipulated. A local attacker who can supply a crafted quantized model could corrupt memory, with possible code execution unknown from the sources.","severity":"medium","cvss":4.8,"epss":null,"exploited":false,"exploited_at":null,"products":[],"ecosystems":[],"tags":["ai-tooling","cuda","exllamav2","gpu","local-inference","python"],"not_affected_if":["models-from-trusted-sources-only","no-exllamav2"],"how_to_check":"- Check the installed version with `pip show exllamav2` and see whether it is `0.3.2` or earlier.\n- Look for any loader that accepts untrusted GPTQ-quantized model files with exllamav2.","fix":"- unknown; the sources say the project has not responded yet. Watch the upstream repository for a patched release and avoid loading untrusted quantized models in the meantime.","status":"open","published_at":"2026-10-26T21:18:12+00:00","first_seen_at":"2026-10-26T21:20:31+00:00","triaged_at":"2026-10-26T21:21:18+00:00","updated_at":"2026-10-26T21:21:18+00:00","category":"ai-tooling","confidence":0.6,"triage":{"at":"2026-10-26T21:21:18Z","model":"claude-haiku-5-5","output":{"fix":"- unknown; the sources say the project has not responded yet. Watch the upstream repository for a patched release and avoid loading untrusted quantized models in the meantime.","reach":"local","title":"ExLlamaV2: crafted g_idx input causes out-of-bounds write locally","impact":"medium","summary":"ExLlamaV2 up to 0.3.2 has an out-of-bounds write in the GPTQ index handling code (QMatrix::make_sequential in the CUDA extension) when the g_idx argument is manipulated. A local attacker who can supply a crafted quantized model could corrupt memory, with possible code execution unknown from the sources.","category":"ai-tooling","nicknames":[],"who_cares":["exllamav2","python","local-inference","gpu","cuda"],"confidence":0.6,"needs_auth":"unknown","worst_case":"local attacker may corrupt memory during model loading","analyst_note":"Confirm whether the out-of-bounds write can be turned into code execution and whether a fixed release exists; check the NVD entry and the upstream ExLlamaV2 repository.","how_to_check":"- Check the installed version with `pip show exllamav2` and see whether it is `0.3.2` or earlier.\n- Look for any loader that accepts untrusted GPTQ-quantized model files with exllamav2.","needs_analyst":true,"not_affected_if":["no-exllamav2","models-from-trusted-sources-only"],"who_is_affected":"Teams that run ExLlamaV2 for local GPU inference on quantized (GPTQ-format) models, especially if they load model files from outside sources. The attack needs local access, so a service that only loads models you built yourself is lower risk. Teams not using ExLlamaV2 can skip it.","user_interaction":"unknown","needs_special_setup":"unknown","severity_if_unknown":"medium"},"cost_usd":0.000657,"input_tokens":3370,"output_tokens":641,"severity_before":"medium"},"analyst_status":null,"analyst_note":null,"search":"'-116502':15A,18A '-2026':14A,17A '0.3.2':32B,93C 'ai':20B 'ai-tool':19B 'argument':57B,118C 'attack':62B,130C 'bound':10A,38B,127C 'caus':6A 'code':45B,75B 'compon':109C 'corrupt':71B 'could':70B 'craft':2A,67B 'cuda':22B,51B 'cve':13A,16A 'earli':143C 'element':96C 'execut':76B,113C 'exllamav2':1A,23B,29B,90C 'exllamav2/exllamav2_ext/cuda/q_matrix.cu':106C 'extens':52B 'file':105C 'flaw':82C 'found':85C 'function':99C 'g':3A,55B,119C 'gptq':42B,110C 'gpu':24B 'handl':44B 'handler':112C 'idx':4A,56B,120C 'impact':95C 'index':43B,111C 'infer':27B 'inform':139C 'input':5A 'issu':146C 'launch':134C 'lead':122C 'local':12A,26B,61B,135C 'local-infer':25B 'make':47B,101C 'manipul':59B,115C 'memori':72B 'model':69B 'need':131C 'org':89C 'out-of-bound':7A,35B,124C 'possibl':74B 'problem':142C 'project':137C 'python':28B 'qmatrix':46B,100C 'quantiz':68B 'report':147C 'respond':151C 'sequenti':48B,102C 'sourc':80B 'suppli':65B 'tool':21B 'turboderp':88C 'turboderp-org':87C 'unknown':77B 'write':11A,39B,128C 'yet':152C","nicknames":[],"mitigations":[],"who_affected":"Teams that run ExLlamaV2 for local GPU inference on quantized (GPTQ-format) models, especially if they load model files from outside sources. The attack needs local access, so a service that only loads models you built yourself is lower risk. Teams not using ExLlamaV2 can skip it.","tier":"B","verified":{},"analyst_revisit_at":null,"first_public_at":"2026-10-26T21:18:12+00:00","kev":null,"cvss_vector":"CVSS:4.0/AV:L/AC:L/AT:N/PR:L/UI:N/VC:N/VI:L/VA:L/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X","cvss_by":"cna@vuldb.com","poc_count":0,"sources":[{"source":"nvd","external_id":"CVE-2026-116502","url":"https://nvd.nist.gov/vuln/detail/CVE-2026-116502","severity":null,"published_at":"2026-10-26T21:18:12+00:00","modified_at":"2026-10-26T21:18:12+00:00","first_seen_at":"2026-10-26T21:20:31+00:00","fetched_at":"2026-10-26T21:20:31+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"}