CVE-2026-85670
Description
A flaw was found in the Hugging Face tokenizers library. A remote attacker could provide a specially crafted tokenizer.json file, which, when processed, leads to an out-of-bounds buffer access. This vulnerability can cause the application to terminate, resulting in a denial of service. In certain configurations, a related defect could also lead to memory corruption.
Statement
Red Hat rates this vulnerability as Important because an attacker can remotely supply a malicious model configuration to trigger an immediate crash of the serving process or potentially compromise memory integrity without needing privileged access. In Red Hat environments hosting AI and machine learning workloads, applications typically ingest tokenizer definitions during model initialization. While exploitation requires the target service to ingest untrusted model artifacts, the resulting disruption can cause severe availability loss across automated model-serving infrastructure.
Understanding the Weakness (CWE)
Integrity
Technical Impact: Modify Memory; Execute Unauthorized Code or Commands
Write operations could cause memory corruption. In some cases, an adversary can modify control data such as return addresses in order to execute unexpected code.
Availability
Technical Impact: DoS: Crash, Exit, or Restart
Attempting to access out-of-range, invalid, or unauthorized memory could cause the product to crash.
Other
Technical Impact: Unexpected State
Subsequent write operations can produce undefined or unexpected results.
Frequently Asked Questions
Not sure what something means? Check out our Security Glossary.
Want to get errata notifications? Sign up here.
