A defensive check in Tesseract is bypassed when a malicious arithmetic model file contains a zero-length stack for NT_SERIES, NT_PARALLEL, or NT_REVERSED layers. The code reaches CacheXScaleFactor and dereferences stack_[0] on an empty vector, invoking an invalid virtual method through a malformed Network object. This deterministic crash occurs during LSTMRecognizer initialization, resulting in a denial of service when the model is loaded. The weakness is an out-of-bounds read and a misuse of object lifetime, reflected in CWE-125 and CWE-824.

The vulnerability affects the tesseract-ocr tesseract engine version 5.5.3 and all earlier releases. Any deployment relying on these versions to load traineddata models is susceptible.

The CVSS score of 6.9 classifies the vulnerability as moderate, though the EPSS score is unavailable, so the exploitation likelihood cannot be precisely quantified. The vulnerability is not listed in the CISA KEV catalog, indicating no known active exploitation. Attackers would need to supply a crafted .traineddata file to the model loader, either locally or through a remote-facing interface that accepts model uploads. The deterministic nature of the crash makes detection straightforward, but the impact can be significant if mission-critical systems rely on continuous OCR processing.