Tesseract OCR CVE-2026-88053

Exploitation requires that the attacker can supply or substitute a .traineddata language model file consumed by the target Tesseract instance - for example, by writing to the tessdata directory, manipulating the TESSDATA_PREFIX environment variable, or controlling any application-level parameter that selects the language data path. The corruption fires during legacy-classifier initialization before OCR processing begins, meaning every Tesseract invocation referencing the malicious file is a trigger event. Exploitation is constrained to standard builds; instances compiled with DISABLED_LEGACY_ENGINE are not affected. Reliable code execution from the heap OOB pointer write is additionally limited by heap layout variability - a crash is the baseline outcome, controlled execution requires heap-shaping effort.

The CVSS 4.0 score of 8.6 with AV:L reflects that an attacker must control a .traineddata file consumed by the target - a prerequisite that substantially limits opportunistic exploitation. The impact metrics VC:H/VI:H/VA:H capture worst-case heap corruption through code execution, though achieving reliable RCE from a heap pointer OOB write requires non-trivial heap layout manipulation and is implementation-dependent. No EPSS data is present in the provided intelligence and this CVE is not listed in CISA KEV, so no confirmed active exploitation exists. The absence of a public exploit further reduces immediate automated exploitation risk. Real-world risk is most elevated in automated document-processing pipelines that ingest .traineddata files from untrusted sources - environments such as document conversion services or ML preprocessing pipelines. Organizations that control their own tessdata directories face meaningfully lower risk. The DISABLED_LEGACY_ENGINE compile-time flag provides a hard mitigation where legacy classifier functionality is not required.

An attacker crafts a .traineddata file containing a TESSDATA_INTTEMP component with NumClassPruners set far above MAX_NUM_CLASS_PRUNERS, then delivers it to a document processing service that passes user-supplied language model paths to Tesseract. On the next OCR invocation, ReadIntTemplates loops past the end of the fixed ClassPruners[] heap array, writing attacker-influenced pointers into adjacent allocations; depending on heap layout, this can yield a crash for denial of service or, with heap grooming, controlled overwrites enabling code execution in the Tesseract process. No public proof-of-concept exploit has been identified at time of analysis.