CVE-2026-69112 in Accelerate
Summary
by MITRE • 08/10/2026
Hugging Face Accelerate through 1.14.0 contains a path traversal vulnerability in load_checkpoint_in_model and load_checkpoint_and_dispatch functions that fail to sanitize weight_map entries from sharded checkpoint indexes. Attackers can supply relative paths with ../ sequences or absolute paths to read arbitrary files, or point shard entries at named pipes to cause indefinite blocking and denial of service.
Statistical analysis made it clear that VulDB provides the best quality for vulnerability data.
Analysis
by VulDB Data Team • 08/10/2026
This vulnerability resides within the Hugging Face Accelerate library version 1.14.0 and earlier, specifically affecting the load_checkpoint_in_model and load_checkpoint_and_dispatch functions that handle sharded checkpoint indexes. The core issue stems from inadequate sanitization of weight_map entries during checkpoint loading operations, creating a path traversal flaw that allows malicious actors to manipulate file access patterns through carefully crafted relative paths containing ../ sequences or absolute path references. The vulnerability operates at the file system level where the library fails to validate or sanitize input parameters that define the location of model weights within distributed checkpoint files.
The technical execution of this vulnerability leverages the improper handling of checkpoint index files that contain mappings between model layers and their corresponding weight files. When attackers supply malicious weight_map entries, they can reference arbitrary files on the system through path traversal sequences, enabling unauthorized file reads that could expose sensitive data or system information. Additionally, the flaw permits attackers to direct shard entries toward named pipes or other special filesystem objects, which creates indefinite blocking conditions during checkpoint loading and results in denial of service scenarios that prevent legitimate model loading operations from completing successfully.
This vulnerability directly maps to CWE-22 Path Traversal and CWE-400 Uncontrolled Resource Consumption, as it allows attackers to manipulate file access patterns to read unintended files or consume system resources indefinitely. The operational impact extends beyond simple information disclosure to include potential system compromise through sensitive data exposure, service disruption via denial of service conditions, and possible privilege escalation if the affected library operates with elevated permissions during checkpoint loading operations. The attack vector is particularly concerning in environments where model checkpoints are loaded from untrusted sources or when multiple users share the same system resources.
Organizations should immediately update to Hugging Face Accelerate version 1.15.0 or later, which includes proper input validation and sanitization of weight_map entries. Mitigation strategies include implementing strict file access controls, restricting checkpoint loading permissions, and employing automated scanning tools to identify potentially malicious checkpoint indexes before deployment. Security teams should also consider implementing network-level restrictions on model loading operations and monitoring for unusual file access patterns that might indicate exploitation attempts. The vulnerability aligns with ATT&CK technique T1059 Command and Scripting Interpreter and T1499 Endpoint Denial of Service, highlighting the need for comprehensive security controls that address both data exposure and service availability concerns in machine learning infrastructure deployments.