{"id":"CVE-2026-114049","kind":"vulnerability","aliases":["CVE-2026-114049"],"related":[],"title":"vLLM: stale token allowlists leak between concurrent requests, constraining other users' outputs","description":"vLLM before 0.31.0 contains a cross-request mask leak in InputBatch.condense() that allows remote attackers to constrain other clients' outputs by leaving stale allowed_token_ids rows. Attackers can submit concurrent constrained generate requests, including via unauthenticated /invocations, so recycled batch slots apply their token allowlist to victims' unrestricted requests.","summary":"vLLM before 0.31.0 fails to clear per-request allowed_token_ids masks when batch slots are recycled in InputBatch.condense(). A remote attacker can make other clients' unrestricted generation requests inherit a token allowlist, constraining their outputs.","severity":"medium","cvss":6.3,"epss":0.00265,"exploited":false,"exploited_at":null,"products":[],"ecosystems":[],"tags":["ai-tooling","llm-serving","python","self-hosted","vllm"],"not_affected_if":["auth-required","no-constrained-generation","single-tenant-deployment"],"how_to_check":"- Check the installed vLLM version with `pip show vllm` or `python -c \"import vllm; print(vllm.__version__)\"`\n- Check whether the server is reachable on an /invocations endpoint without authentication, which the sources say is exploitable","fix":"- Upgrade vLLM to `0.31.0` or later\n- Until upgraded, restrict access to the /invocations endpoint behind authentication","status":"open","published_at":"2026-10-21T01:17:27+00:00","first_seen_at":"2026-10-23T21:57:27+00:00","triaged_at":"2026-10-25T04:01:25+00:00","updated_at":"2026-10-25T04:01:25+00:00","category":"ai-tooling","confidence":0.7,"triage":{"at":"2026-10-25T04:01:25Z","model":"claude-haiku-5-5","output":{"fix":"- Upgrade vLLM to `0.31.0` or later\n- Until upgraded, restrict access to the /invocations endpoint behind authentication","reach":"internet","title":"vLLM: stale token allowlists leak between concurrent requests, constraining other users' outputs","impact":"medium","summary":"vLLM before 0.31.0 fails to clear per-request allowed_token_ids masks when batch slots are recycled in InputBatch.condense(). A remote attacker can make other clients' unrestricted generation requests inherit a token allowlist, constraining their outputs.","category":"ai-tooling","nicknames":[],"who_cares":["vllm","self-hosted","llm-serving","python"],"confidence":0.7,"needs_auth":"no","worst_case":"other clients' outputs are constrained by another request's token allowlist","analyst_note":"","how_to_check":"- Check the installed vLLM version with `pip show vllm` or `python -c \"import vllm; print(vllm.__version__)\"`\n- Check whether the server is reachable on an /invocations endpoint without authentication, which the sources say is exploitable","needs_analyst":false,"not_affected_if":["single-tenant-deployment","auth-required","no-constrained-generation"],"who_is_affected":"Teams running vLLM as a shared inference server that handles concurrent constrained generation requests from multiple clients, especially where /invocations is exposed without authentication. Single-user or isolated deployments that never mix constrained and unconstrained requests are much less exposed.","user_interaction":"no","needs_special_setup":"yes","severity_if_unknown":"medium"},"cost_usd":0.00062,"input_tokens":3392,"output_tokens":561,"severity_before":"medium"},"analyst_status":null,"analyst_note":null,"search":"'-114049':15A,18A '-2026':14A,17A '/invocations':105C '0.31.0':32B,69C 'ai':20B 'ai-tool':19B 'allow':39B,80C,91C 'allowlist':4A,63B,113C 'appli':110C 'attack':52B,82C,95C 'batch':44B,108C 'clear':35B 'client':56B,86C 'concurr':7A,98C 'constrain':9A,64B,84C,99C 'contain':70C 'cross':73C 'cross-request':72C 'cve':13A,16A 'fail':33B 'generat':58B,100C 'host':28B 'id':41B,93C 'includ':102C 'inherit':60B 'inputbatch.condense':49B,78C 'leak':5A,76C 'leav':89C 'llm':23B 'llm-serv':22B 'make':54B 'mask':42B,75C 'output':12A,66B,87C 'per':37B 'per-request':36B 'python':25B 'recycl':47B,107C 'remot':51B,81C 'request':8A,38B,59B,74C,101C,117C 'row':94C 'self':27B 'self-host':26B 'serv':24B 'slot':45B,109C 'stale':2A,90C 'submit':97C 'token':3A,40B,62B,92C,112C 'tool':21B 'unauthent':104C 'unrestrict':57B,116C 'user':11A 'via':103C 'victim':115C 'vllm':1A,29B,30B,67C","nicknames":[],"mitigations":[],"who_affected":"Teams running vLLM as a shared inference server that handles concurrent constrained generation requests from multiple clients, especially where /invocations is exposed without authentication. Single-user or isolated deployments that never mix constrained and unconstrained requests are much less exposed.","tier":"B","verified":{},"analyst_revisit_at":null,"first_public_at":"2026-10-21T01:17:27+00:00","kev":null,"cvss_vector":"CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/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":"disclosure@vulncheck.com","poc_count":0,"sources":[{"source":"nvd","external_id":"CVE-2026-114049","url":"https://nvd.nist.gov/vuln/detail/CVE-2026-114049","severity":null,"published_at":"2026-10-21T01:17:27+00:00","modified_at":"2026-10-22T15:44:59+00:00","first_seen_at":"2026-10-23T21:57:27+00:00","fetched_at":"2026-10-23T22:29:01+00:00","note":null},{"source":"epss","external_id":"CVE-2026-114049","url":"https://api.first.org/data/v1/epss?cve=CVE-2026-114049","severity":null,"published_at":null,"modified_at":null,"first_seen_at":"2026-10-24T13:15:55+00:00","fetched_at":"2026-10-24T13:15:55+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"}