CVE-2026-115533

Untrusted LeRobot policy configurations could execute code during loading

Hugging Face · LeRobot policy processor loading

First documented disclosure · Review checkpoint

Summary

Loading a policy from an untrusted repository could run code with the privileges of the LeRobot process.

Severity and exploitation evidence

High · 8.8/10 · CVSS 3.1. Severity source.

Exploitation not established: A researcher demonstrated the behavior in controlled tests. CISA’s November 16 assessment recorded no known exploitation.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The processor loader accepted class names from policy configuration and instantiated them without limiting them to trusted processor types.

Conditions required

An operator or application loads an untrusted policy configuration on an affected installation.

Potential impact

A malicious policy author could gain code execution within the policy-loading process.

Affected versions and configurations

0.4.0 through 0.6.1 according to the CNA.

Fix and mitigation status

No verified fixed version. The reviewed 0.7.0 release notes do not identify this fix.

Practical next steps

Use policies from trusted sources, pin reviewed revisions and isolate the policy-loading process from credentials and valuable files.

Operating lesson

Treat policy processor configuration as executable input unless its allowed types are explicitly constrained.

Evidence and limits

Project issue 4219 publicly described the flaw on July 29. The October 26 CVE lists affected releases. On November 16 CISA added CVSS 3.1 High, 8.8, to the record. The proposed upstream correction remained open at review. The existing LeRobot CVE-2026-25874 concerns a different network message handler.

Sources