CVE-2026-88049

Tesseract OCR versions 5.5.3 and earlier contain a heap out-of-bounds write (CWE-787) in the LSTM recognition path: while earlier hardening added bounds checks to NetworkIO::CopyTimeStepGeneral and NetworkIO::Randomize, the NetworkIO::WriteTimeStepPart and NetworkIO::AddTimeStepPart functions in src/lstm/networkio.cpp remained unchecked. The flaw is triggered when the default LSTM engine processes a specially crafted .traineddata model file: in LSTM::Forward (src/lstm/lstm.cpp) the source_ buffer is sized from the independently deserialized na_ field, but WriteTimeStepPart writes a count driven by ns_ from the CI gate WeightMatrix dim1() value, so a crafted NT_LSTM layer with ns_ much larger than na_ overruns the heap buffer during the first recognition step. An attacker who can get such a model file processed gains heap corruption, causing a crash or potentially controlled memory corruption; CVSS 4.0 rates this 8.6 High with a local attack vector and no privileges or user interaction required. Anyone running Tesseract 5.5.3 or earlier is affected in principle, with practical risk concentrated in applications and pipelines that load third-party or otherwise untrusted .traineddata language models rather than only official trusted ones. No public proof-of-concept or in-the-wild exploitation is known, the flaw is not in CISA KEV, and no fixed release was available at the time of review.

What to do: No fixed Tesseract release is available yet, so monitor the project for a patched version and upgrade as soon as one ships. Until then, avoid loading .traineddata files from untrusted or third-party sources on affected builds, verify the provenance and integrity of any externally sourced language models, and consider running OCR of external models in sandboxed, low-privilege workers to limit impact of a crash or heap corruption.