Code Injection in Flash-Attention Training Script

Executive Summary

This vulnerability exists in the flash-attention project's training script as of commit e724e2588cbe754beb97cf7c011b5e7e34119e62. The script registers Python's eval() function as a Hydra configuration resolver named eval. This means that configuration files can use the ${eval:...} syntax to execute arbitrary Python code.

An attacker can exploit this by supplying a malicious configuration file that contains harmful Python code. When the training script runs with this configuration, it will execute the attacker's code, leading to arbitrary code execution.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can lead to arbitrary code execution on the system running the flash-attention training script. An attacker who provides a malicious configuration file can execute any Python code, potentially compromising the system, stealing data, modifying files, or disrupting operations.

Useful 0 Not Useful 0

Mitigation Strategies

To mitigate this vulnerability, avoid using untrusted or malicious configuration files with the flash-attention training script, as these can execute arbitrary Python code via the eval resolver.

Consider removing or disabling the registration of the Python eval() function as a Hydra configuration resolver to prevent execution of arbitrary code through the ${eval:...} syntax.

Ensure that only trusted users have access to modify or provide configuration files used by the training script.

Useful 0 Not Useful 0