AI Stack Alert Medium CVE-2026-126395 For CISO, SOC / IR, AppSec, AI Eng AI-assisted
CVE-2026-126395: vLLM before 0.31.0 leaks heap addresses via uppercase PIL image address errors
vLLM from 0.14.0rc1 before 0.31.0 has an information disclosure flaw in sanitize_message(). Its lowercase-only regex does not scrub uppercase PIL image addresses from error messages. An unauthenticated attacker can send a crafted PNG to endpoints such as /tokenize to trigger leaked live server heap addresses.
Why it matters
Any exposed vLLM server before 0.31.0 can be probed without credentials. The leaked heap addresses can help an attacker bypass ASLR, which makes follow-on memory corruption attacks easier.
Do this
- Upgrade vLLM to 0.31.0 or later on all inference servers.
- Restrict network access to vLLM endpoints such as /tokenize to trusted clients until patched.
- Review vLLM logs for unauthenticated requests to /tokenize that include image uploads and errors from the HF processor.
- Confirm sanitize_message() behavior in patched builds with a test that uses non-lowercase PIL image addresses in a safe, non-exploit test case.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 6.9CVSS
- NoPublic PoC
- 27/100Priority
- 80%Confidence
Evidence
“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.”
Sources
Affected / related
vLLM, vllm
Published 2026-11-13 10:17 UTC · updated 2026-11-13 12:35 UTC · AI-assisted summary, checked against quoted sources · How we score