Vulnary
← back to the feed
Critical· 9.6

CVE-2026-5241

A flaw in huggingface transformers 5.2.0 lets an attacker-controlled model repository run arbitrary code when a LightGlue model is loaded, even if the trust_remote_code flag is set to False. The bug occurs because the flag is overridden by untrusted configuration data in a nested code path. This can allow attackers to execute code on the host system.

publishedJun 3, 2026
last modifiedJul 22, 2026
sourcesNVD
severity · cvss
9.6
critical · how bad it is
exploitation · epss
<1%
43th percentile · chance of exploitation in 30 days
(ai-assisted) A model wrote this summary from the official data, so double-check it against the source before you act on it. Read the official advisory →
auto-deletes from the system
counting…on Sep 4, 2026

No official fix yet. If none appears within 45 days of first tracking, this entry is removed automatically.

01

Who is affected

huggingface transformers 5.2.0, commonly used in API inference servers, research notebooks, CI/CD pipelines, and model evaluation workers.

02

Real-world impact

An attacker could run arbitrary code on the system that loads the model, potentially stealing credentials, moving laterally, or installing a persistent backdoor.

03

Why this severity

The CVSS score of 9.6 reflects the high impact on confidentiality, integrity, and availability, the fact that the vulnerability can be exploited remotely without authentication, and that it requires only low attack complexity and minimal user interaction.

04

What to do about it

no official fix yet
interim mitigations
  • Avoid loading models from untrusted repositories or sources.
  • Use only trusted model repositories and verify the model’s integrity before loading.

No official fix or mitigation is documented in the sources yet. Monitor the vendor advisory and apply the patch as soon as it is released.

NVD description

05

Timeline

  1. Jun 3, 2026 · Jun 3, 2026
    Published
    Disclosed and added to the National Vulnerability Database.
  2. Jul 22, 2026 · 13d ago
    Advisory updated
    The NVD record was last revised.
06

How it’s attacked

Attack vectorNetwork (remote)
Attack complexityLow
Privileges requiredNone
User interactionRequired
ScopeChanged
Confidentiality impactHigh
Integrity impactHigh
Availability impactHigh
07

References & advisories

(ai-assisted) A model wrote this summary from the official data, so double-check it against the source before you act on it. Read the official advisory →