vLLM: Remote code execution via request-controlled code_revision
Summary
vLLM from 0.7.3 before 0.31.0 passes the request parameter mm_processor_kwargs.code_revision to the Transformers AutoProcessor loader. If trust_remote_code is enabled and a multimodal model with dynamic processor code is served, an attacker who controls a revision of the model repository can execute arbitrary Python in the vLLM API process. The unauthenticated /tokenize endpoint can be used for this, even when an API key is configured.
Affected
- vLLM from 0.7.3 before 0.31.0
- Only if trust_remote_code is enabled and the served model uses dynamic AutoProcessor code
- Default configuration (trust_remote_code=False) is not affected
Fixed in
- Fixed in vLLM 0.31.0 (upstream fix commit 570735520942b6428ae98f1eb442c6d88adae55e)
- Note: a separate RCE, GHSA-h3rc-j82p-2c3h (CVE-2026-115915), is fixed only in 0.32.0
Workaround
- Keep trust_remote_code disabled
- Do not expose the vLLM API to untrusted networks
- Restrict access to /tokenize with an authenticating reverse proxy or firewall
- Unknown: no vendor-documented workaround for unpatched systems beyond these
Exploitation
- No confirmed known exploitation found (not listed as known exploited)
- No public proof of concept found in searches; this does not rule one out
- CISA SSVC: exploitation none, automatable no, technical impact total
Our recommendation
- Upgrade vLLM to 0.31.0 or later, and to 0.32.0 or later to also cover CVE-2026-115915
- Check whether trust_remote_code is enabled and whether multimodal models are served
- Verify that the /tokenize endpoint is not reachable from the internet
- Review that per-request mm_processor_kwargs usage still works after the upgrade, as 0.31.0 may reject it unless configured
- Monitor logs for unusual requests to /tokenize
Public exploits
PoC-in-GitHub, Metasploit and Exploit-DB list nothing for this CVE.
Research sources
- GitHub Security Advisory GHSA-h3rc-6mm3-gc2m github.com
- Fix commit 570735520942b6428ae98f1eb442c6d88adae55e github.com
- VulnCheck advisory www.vulncheck.com
- OSV GHSA-h3rc-6mm3-gc2m osv.dev
- OffSeq threat radar entry radar.offseq.com
- Hol.org CVE summary hol.org
- LinuxSecurity news linuxsecurity.com
- OpenCVE CVE-2026-114053 app.opencve.io
Researched and written up in English and German by an AI (Claude by Anthropic) based on NVD, CISA, EPSS and the sources listed. The vendor information is authoritative. Researched on 31 Oct 2026, 04:11.
Original description (NVD)
vLLM from 0.7.3 before 0.31.0 forwards the request-supplied mm_processor_kwargs.code_revision to Transformers AutoProcessor.from_pretrained, letting clients select which revision of a model's dynamic processor code is imported when trust_remote_code is enabled. An attacker controlling a revision in the served model repository can send a multimodal request to the unauthenticated /tokenize endpoint to execute arbitrary Python in the vLLM API process.
Sources
The tracker combines official, freely usable data sources – updated continuously:
- NVD (NIST) – CVEs, rating (CVSS), affected products
- CISA KEV – vulnerabilities confirmed by CISA as actively exploited
- CISA Vulnrichment – CISA assessment (SSVC) in the official CVE record
- FIRST EPSS – probability of exploitation
- Microsoft MSRC – Patch Tuesday: vulnerabilities, severity, updates
- PoC-in-GitHub – public proof-of-concept repositories
- Metasploit (Rapid7) – modules of the Metasploit Framework
- Exploit-DB (OffSec) – archive of public exploits (links only)
- Patch days: official vendor pages (linked per vendor) (Vendor patch days)
This product uses data from the NVD API but is not endorsed or certified by the NVD. EPSS data © FIRST.org.
Affected? We can help.
We check whether your systems are affected, install updates and close the gap – continuously with the System Health Checker if you wish.