CVE-2026-31253
HighSummary
The flash-attention training framework contains an insecure deserialization vulnerability in its checkpoint loading mechanism. The load_checkpoint() function and the checkpoint loading code in eval.py use torch.load() without enabling the weights_only=True parameter, allowing the deserialization of arbitrary Python objects.
Risk Assessment
An attacker can exploit this vulnerability by providing a maliciously crafted checkpoint file, leading to the execution of arbitrary code on the victim's system. This poses a significant security risk.
Recommendation
It is recommended to enable the weights_only=True parameter in the torch.load() function to minimize risk. Additionally, verify the source of checkpoint files before loading them.
Related vulnerabilities
- CVE-2026-8778Critical
The MIPL Grouped Checkout Fields for WooCommerce plugin for WordPress (versions up to and including 1.2.1) is vulnerable to arbitrary file uploads due to missing file type validation in the `mipl_wc_upload_file` function. This allows unauthenticated attackers to upload arbitrary files to the server, potentially enabling remote code execution.
- CVE-2026-82107Critical
IBM DataStage on Cloud Pak for Data 5.4.0.0 contains an improper authentication flaw that allows a remote authenticated attacker to obtain sensitive information and bypass security restrictions.
- CVE-2026-82100Critical
IBM DataStage on Cloud Pak for Data 5.4.0.0 is vulnerable to a path traversal issue that allows a remote authenticated attacker to cause a denial of service.
- CVE-2026-81204Critical
IBM Langflow OSS versions 1.0.0 through 1.11.5 allow a remote attacker to execute arbitrary code due to code injection during graph construction.
- CVE-2026-80424Critical
IBM DataStage on Cloud Pak for Data 5.4.0.0 allows a remote authenticated attacker to create arbitrary files due to path traversal during archive extraction.
- CVE-2026-79724Critical
IBM Langflow OSS versions 1.0.0 through 1.11.5 allow a remote attacker to execute arbitrary OS commands due to improper neutralization of special elements used in an OS command.
- CVE-2026-78573Critical
IBM ContextForge MCP Gateway versions 1.0.0 through 1.0.7 allow a remote attacker to gain administrative access due to the use of default credentials.
- CVE-2026-45764Critical
Suricata is a network Intrusion Detection System, Intrusion Prevention System and Network Security Monitoring engine. Prior to versions 7.0.16 and 8.0.5, a protocol change while processing HTTP/2 traffic could lead to type confusion, causing Suricata to crash and resulting in denial of service.
- CVE-2026-19646Critical
IBM Common Licensing Agent 9.0, 9.0.0.1, 9.0.0.2 and ART 9.0, 9.0.0.1, 9.0.0.2 allow a remote attacker to redirect users to an arbitrary domain due to improper validation of the HTTP Host header.
- CVE-2026-89094Critical
Forgejo before 16.0.4 allows remote code execution via a crafted template repository because template expansion on files in .forgejo/template is mishandled.
Original NVD description (English source)
The flash-attention training framework thru commit e724e2588cbe754beb97cf7c011b5e7e34119e62 (2025-13-04) contains an insecure deserialization vulnerability (CWE-502) in its checkpoint loading mechanism. The load_checkpoint() function in checkpoint.py and the checkpoint loading code in eval.py use torch.load() without enabling the security-restrictive weights_only=True parameter. This allows the deserialization of arbitrary Python objects via the pickle module. An attacker can exploit this by providing a maliciously crafted checkpoint file. When a victim loads this checkpoint during model warmstarting or evaluation, arbitrary code is executed on the victim's system.