Hugging Face Transformers SEW-D convert_config コードインジェクションによるリモートコード実行脆弱性

The impact of this vulnerability is significant, as it allows an attacker to gain complete control over the affected system. By crafting a malicious checkpoint and tricking a user into converting it, an attacker can execute arbitrary code in the context of the Transformers process. This could lead to data theft, system compromise, and potentially lateral movement within the network. The vulnerability's reliance on user interaction to convert the malicious checkpoint limits the immediate exploitability, but the potential for widespread impact remains high, especially in environments where checkpoint conversion is a common operation.

Organizations heavily reliant on Hugging Face Transformers for natural language processing tasks, particularly those involved in model training and deployment, are at significant risk. Environments where checkpoint conversion is automated or performed by less experienced users are especially vulnerable. Shared hosting environments where multiple users have access to the Transformers installation are also at increased risk.

• python / transformers:

import subprocess
# Check for the vulnerable version
result = subprocess.run(['pip', 'show', 'transformers'], capture_output=True, text=True)
if 'Version: 4.57.0' in result.stdout:
    print('Vulnerable version detected!')

• python / transformers: Monitor for unusual process execution after checkpoint conversion using system monitoring tools. • generic web: Monitor access logs for requests related to checkpoint conversion endpoints. Look for unusual user agents or request parameters.

1. 2025-12-23Disclosure

   disclosure

1. 予約済み2025-12-18
2. 公開日2025-12-23
3. 更新日2025-12-24
4. EPSS 更新日2026-03-23

The primary mitigation for CVE-2025-14927 is to upgrade Hugging Face Transformers to version 4.57.1 or later. Until the upgrade can be performed, consider restricting access to checkpoint conversion functionality to trusted users only. Implement input validation and sanitization on any user-supplied data used in the conversion process. While a direct WAF rule is unlikely to be effective, monitoring for unusual process execution patterns after checkpoint conversion could provide early detection. After upgrading, confirm the fix by attempting to convert a known-safe checkpoint and verifying that no unexpected code is executed.