The vulnerability lies in the load_ckpt function of torchdistill’s common/main_util.py, where a manipulated src_ckpt argument causes unsafe deserialization. This flaw, classified as CWE-20 (Improper Input Validation) and CWE-502 (Deserialization of Untrusted Data), permits an attacker who can influence the parameter to load maliciously crafted data, leading to code execution or other unintended behavior. The impact is limited to the environment where the function is executed and requires the attacker to have local access to invoke the function.

The flaw affects the open-source library torchdistill maintained by yoshitomo-matsubara, specifically versions up to and including 1.1.5. users of these versions that import and use load_ckpt should consider themselves vulnerable. No specific operating systems or platforms are singled out; the issue depends on local execution of the Python code.

The CVSS score of 4.8 indicates a low severity risk. EPSS data is not available, so the likelihood of exploitation is unclear, but the fact that an exploit has been published and the flaw requires local execution implies that safety is only relevant if the code is run in an environment with untrusted input. The vulnerability is not listed in the CISA KEV catalog. Attackers would need to gain local access to run load_ckpt with a crafted checkpoint, often common in research or deployment scenarios where checkpoint files are loaded from untrusted sources.