The flash-attention project, as of commit e724e2588cbe754beb97cf7c011b5e7e34119e62 dated 2025-13-04, contains a code injection vulnerability (CWE-94) in its training script. This occurs because the script exposes the Python eval() function as a Hydra configuration resolver under the name 'eval'. This design permits configuration files to execute arbitrary Python code via the ${eval:...} syntax. An attacker able to provide a malicious configuration file can exploit this to execute arbitrary code during the training process.

Successful exploitation allows arbitrary code execution on the system running the training script, potentially leading to full compromise of the environment where the script is executed. This could result in unauthorized actions, data manipulation, or system control depending on the privileges of the executing process.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is available, avoid running the training script with untrusted or unauthenticated configuration files to prevent exploitation. Restrict access to configuration files and validate or sanitize inputs where possible.