Skip to main content
Public proof of conceptFix available

CVE-2026-54236

A public proof of concept exists; no exploitation attested.

  • Published2026-06-22
  • Fixed in0.24.0
  • WeaknessCWE-532

Am I affected?

With a domain, we read the version its public headers and pages expose. A distribution build can carry the fix under an older number.

Description

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, the fix for CVE-2026-22778, which introduced a sanitize_message helper that strips object-repr memory addresses from error messages before they reach the client, is incomplete: several response paths echo str(exc) directly to clients without calling sanitize_message. The unsanitized sites include the Anthropic API router in vllm/entrypoints/anthropic/api_router.py (the POST /v1/messages and POST /v1/messages/count_tokens handlers), the Server-Sent Events streaming converter in vllm/entrypoints/anthropic/serving.py, and the realtime speech-to-text WebSocket in vllm/entrypoints/speech_to_text/realtime/connection.py. These paths catch the exception inside the route coroutine and construct the JSONResponse themselves, bypassing the sanitizing global FastAPI exception handler, and WebSocket frames do not traverse that handler chain at all. Using the same primitive as the parent issue, an unauthenticated attacker can send malformed image bytes through the Anthropic Messages API image content parts so that PIL.Image.open raises an UnidentifiedImageError whose message contains the BytesIO object repr, leaking the heap memory address verbatim in the error.message field of the response body. This vulnerability is fixed in 0.23.1rc0.

Timeline

dates in UTC
  1. 2026-06-2223:16CVE publishedNVD

Observed activity

isMalicious corpus · updated 2026-10-23 21:37 UTC
Indicators tagged with this CVE, per day0 indicators over 7 days
Block these indicators
  • Firewall list (EDL, text): https://api.ismalicious.com/edl/cve/CVE-2026-54236?format=txt
  • CSV export: https://api.ismalicious.com/edl/cve/CVE-2026-54236?format=csv
  • TAXII collection of exploited CVEs
  • The EDL and CSV take an API key on the Pro plan.

Versions per branch

OSV · NVD
vllm
BranchAffectedFixed in
0.24< 0.24.00.24.0
0.23≤ 0.23.0no fix published
Vllm
BranchAffectedFixed in
0.23< 0.23.10.23.1

as publishedRanges as NVD and OSV publish them; a distribution’s build can be fixed under an older number.

Fix

  1. Update vllm to 0.24.0

Advisories and patches

Detection

Nuclei template
none so far
ProjectDiscovery
Exploit-DB
no entry
Exploit-DB
Proof-of-concept repositories
none listed
PoC-in-GitHub

Sources

  • NVDCVE recordread 2026-10-23
  • CISA-ADPSSVC decisionpoc2026-06-23
  • OpenCVESSVC decisionpoc—
  • CNACVSS 3.15.3read 2026-10-23
  • CNAweakness (CWE)CWE-532read 2026-10-23
  • FIRST EPSSEPSS score0.9 %2026-10-22
  • OSVaffected packages—

Technical details

CVSS v3 Vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:N/A:N
Published
2026-06-22
Last Modified
2026-06-24
NVD status
Analyzed

Frequently asked questions

What is CVE-2026-54236?

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, the fix for CVE-2026-22778, which introduced a sanitize_message helper that strips object-repr memory addresses from error messages before they reach the client, is incomplete: several response paths echo str(exc) directly to clients without calling sanitize_message. The unsanitized sites include the Anthropic API router in vllm/entrypoints/anthropic/api_router.py (the POST /v1/messages and POST /v1/messages/count_tokens handlers), the Server-Sent Events streaming converter in vllm/entrypoints/anthropic/serving.py, and the realtime speech-to-text WebSocket in vllm/entrypoints/speech_to_text/realtime/connection.py. These paths catch the exception inside the route coroutine and construct the JSONResponse themselves, bypassing the sanitizing global FastAPI exception handler, and WebSocket frames do not traverse that handler chain at all. Using the same primitive as the parent issue, an unauthenticated attacker can send malformed image bytes through the Anthropic Messages API image content parts so that PIL.Image.open raises an UnidentifiedImageError whose message contains the BytesIO object repr, leaking the heap memory address verbatim in the error.message field of the response body. This vulnerability is fixed in 0.23.1rc0.

Is CVE-2026-54236 actively exploited?

A proof-of-concept exploit exists for CVE-2026-54236, but active exploitation has not been confirmed at this time.

What is the CVSS score for CVE-2026-54236?

CVE-2026-54236 has a CVSS v3 base score of 5.3 (MEDIUM severity), with vector string CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:N/A:N.

Running vllm? Hear about its next vulnerability

Surveillance reads the versions in your stack — a domain, a lockfile or a list — checks each CVE against the ranges its sources publish, and alerts you when one you run is exploited or enters CISA KEV. A version it cannot read stays « version not verified ».

Get alerts for this productCISA KEV additions, October 2026Search CVEs

No credit card required · No monthly cap on website lookups · 250 API requests/month

Go further

Related vulnerabilities

Same product first, then the same vendor advisory and the same publication week.

Back to CVE Database