{"id":"CVE-2026-126395","kind":"vulnerability","aliases":["CVE-2026-126395"],"related":[],"title":"vLLM: error messages leak heap addresses, helping attackers bypass memory protection","description":"vLLM from 0.14.0rc1 before 0.31.0 contains an information disclosure vulnerability in sanitize_message() whose lowercase-only regex fails to scrub uppercase PIL image addresses. Unauthenticated attackers can send a crafted PNG to endpoints such as /tokenize, triggering HF-processor errors that leak live server heap addresses useful for ASLR bypass.","summary":"vLLM's error sanitizer misses uppercase memory addresses in image-processing errors, so a crafted PNG sent to endpoints such as /tokenize can make the server reveal live heap addresses. An attacker can use those addresses to defeat address space layout randomization, which is a step toward a larger memory-corruption attack.","severity":"medium","cvss":6.9,"epss":null,"exploited":false,"exploited_at":null,"products":[],"ecosystems":[],"tags":["ai-inference","ai-tooling","python","self-hosted","vllm"],"not_affected_if":["default-config","private-network-only"],"how_to_check":"- Check the installed vLLM version with `pip show vllm` and compare it with the fixed version in the product ranges.\n- Look for exposed vLLM HTTP endpoints such as `/tokenize` reachable from untrusted networks.","fix":"- vLLM: upgrade to `0.31.0` or later, per the affected range before `0.31.0`.","status":"open","published_at":"2026-11-13T10:17:43+00:00","first_seen_at":"2026-11-13T10:47:18+00:00","triaged_at":"2026-11-13T10:54:47+00:00","updated_at":"2026-11-13T10:54:47+00:00","category":"ai-tooling","confidence":0.7,"triage":{"at":"2026-11-13T10:54:47Z","model":"claude-haiku-5-5","output":{"fix":"- vLLM: upgrade to `0.31.0` or later, per the affected range before `0.31.0`.","reach":"internet","title":"vLLM: error messages leak heap addresses, helping attackers bypass memory protection","impact":"low","summary":"vLLM's error sanitizer misses uppercase memory addresses in image-processing errors, so a crafted PNG sent to endpoints such as /tokenize can make the server reveal live heap addresses. An attacker can use those addresses to defeat address space layout randomization, which is a step toward a larger memory-corruption attack.","audience":"developers","category":"ai-tooling","nicknames":[],"who_cares":["vllm","python","ai-inference","self-hosted"],"confidence":0.7,"fix_status":"available","needs_auth":"no","worst_case":"Server heap addresses leak, aiding a follow-on memory exploit","analyst_note":"Confirm the endpoint list and any special configuration needed for the leak, using the maintainer advisory on github.com and the VulnCheck advisory.","how_to_check":"- Check the installed vLLM version with `pip show vllm` and compare it with the fixed version in the product ranges.\n- Look for exposed vLLM HTTP endpoints such as `/tokenize` reachable from untrusted networks.","plain_action":"","product_name":"","needs_analyst":true,"not_affected_if":["private-network-only","default-config"],"who_is_affected":"Teams running vLLM from 0.14.0rc1 up to but not including 0.31.0 that expose its HTTP API, especially endpoints that accept image input such as /tokenize. Deployments that keep vLLM on a private network with no untrusted callers are lower risk. unknown for further detail beyond the sources.","user_interaction":"no","product_ecosystem":"","needs_special_setup":"unknown","severity_if_unknown":"medium"},"cost_usd":0.000788,"input_tokens":4639,"output_tokens":649,"severity_before":"medium"},"analyst_status":null,"analyst_note":null,"search":"'-126395':14A,17A '-2026':13A,16A '/tokenize':51B,120C '0.14.0':85C '0.31.0':88C 'address':6A,36B,59B,65B,68B,108C,131C 'ai':19B,22B 'ai-infer':18B 'ai-tool':21B 'aslr':134C 'attack':8A,61B,82B,110C 'bypass':9A,135C 'contain':89C 'corrupt':81B 'craft':44B,114C 'cve':12A,15A 'defeat':67B 'disclosur':92C 'endpoint':48B,117C 'error':2A,31B,41B,125C 'fail':102C 'heap':5A,58B,130C 'help':7A 'hf':123C 'hf-processor':122C 'host':27B 'imag':39B,107C 'image-process':38B 'infer':20B 'inform':91C 'larger':78B 'layout':70B 'leak':4A,127C 'live':57B,128C 'lowercas':99C 'lowercase-on':98C 'make':53B 'memori':10A,35B,80B 'memory-corrupt':79B 'messag':3A,96C 'miss':33B 'pil':106C 'png':45B,115C 'process':40B 'processor':124C 'protect':11A 'python':24B 'random':71B 'rc1':86C 'regex':101C 'reveal':56B 'sanit':32B,95C 'scrub':104C 'self':26B 'self-host':25B 'send':112C 'sent':46B 'server':55B,129C 'space':69B 'step':75B 'tool':23B 'toward':76B 'trigger':121C 'unauthent':109C 'uppercas':34B,105C 'use':63B,132C 'vllm':1A,28B,29B,83C 'vulner':93C 'whose':97C","nicknames":[],"mitigations":[],"who_affected":"Teams running vLLM from 0.14.0rc1 up to but not including 0.31.0 that expose its HTTP API, especially endpoints that accept image input such as /tokenize. Deployments that keep vLLM on a private network with no untrusted callers are lower risk. unknown for further detail beyond the sources.","tier":"B","verified":{},"analyst_revisit_at":null,"fix_status":"available","audience":"developers","plain_action":null,"pkeys":[],"ptokens":[],"labels":["aiinference","aitooling","medium","python","selfhosted","vllm"],"campaign_type":null,"search_a":"'-126395':14A,17A '-2026':13A,16A 'address':6A 'attack':8A 'bypass':9A 'cve':12A,15A 'error':2A 'heap':5A 'help':7A 'leak':4A 'memori':10A 'messag':3A 'protect':11A 'vllm':1A","search_b":"'/tokenize':51B 'address':36B,59B,65B,68B 'ai':19B,22B 'ai-infer':18B 'ai-tool':21B 'attack':61B,82B 'corrupt':81B 'craft':44B 'defeat':67B 'endpoint':48B 'error':31B,41B 'heap':58B 'host':27B 'imag':39B 'image-process':38B 'infer':20B 'larger':78B 'layout':70B 'live':57B 'make':53B 'memori':35B,80B 'memory-corrupt':79B 'miss':33B 'png':45B 'process':40B 'python':24B 'random':71B 'reveal':56B 'sanit':32B 'self':26B 'self-host':25B 'sent':46B 'server':55B 'space':69B 'step':75B 'tool':23B 'toward':76B 'uppercas':34B 'use':63B 'vllm':28B,29B","search_c":"'/tokenize':120C '0.14.0':85C '0.31.0':88C 'address':108C,131C 'aslr':134C 'attack':110C 'bypass':135C 'contain':89C 'craft':114C 'disclosur':92C 'endpoint':117C 'error':125C 'fail':102C 'heap':130C 'hf':123C 'hf-processor':122C 'imag':107C 'inform':91C 'leak':127C 'live':128C 'lowercas':99C 'lowercase-on':98C 'messag':96C 'pil':106C 'png':115C 'processor':124C 'rc1':86C 'regex':101C 'sanit':95C 'scrub':104C 'send':112C 'server':129C 'trigger':121C 'unauthent':109C 'uppercas':105C 'use':132C 'vllm':83C 'vulner':93C 'whose':97C","cause_entry":null,"cause_path":null,"disclosures_checked_at":null,"first_public_at":"2026-10-27T22:14:15+00:00","kev":null,"cvss_vector":"CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:L/VI:N/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-2026-126395","url":"https://nvd.nist.gov/vuln/detail/CVE-2026-126395","severity":null,"published_at":"2026-11-13T10:17:43+00:00","modified_at":"2026-11-13T10:17:43+00:00","first_seen_at":"2026-11-13T10:47:18+00:00","fetched_at":"2026-11-13T10:47:18+00:00","note":null}],"documents":[{"id":110587,"url":"https://github.com/vllm-project/vllm/security/advisories/GHSA-m9fg-w9qh-67pc","title":"Maintainer's advisory: github.com","kind":"vendor","source":"ghsa-repo","publisher":"vendor","text":null,"published_at":"2026-10-27T22:14:15+00:00","relation":"advisory","confidence":null,"note":"Vendor advisory: github.com"},{"id":145033,"url":"https://www.vulncheck.com/advisories/vllm-0.14-0rc1-before-0.31.0-heap-address-disclosure-via-sanitize-message","title":"Third-party advisory: www.vulncheck.com","kind":"advisory","source":"nvd","publisher":"research","text":null,"published_at":null,"relation":"advisory","confidence":null,"note":"Third-party advisory: www.vulncheck.com"}],"disclosures":[],"exploit_sources":[],"notice":"Free to read, with fair use of about 60 pages a minute. For higher limits or API access: sales@infrastack.ai"}