The vulnerability arises in the SymSpellModelSerializer.create() method of Apache OpenNLP's opennlp-spellcheck extension introduced in version 3.0.0-M4. It reads two 32-bit signed integer count fields from a binary SymSpell model stream and uses these values to pre-size LinkedHashMaps without an upper bound, allowing an attacker to specify extremely large counts. This causes the JVM to allocate very large backing arrays (4–8 GB), resulting in an OutOfMemoryError and denial of service. The flaw affects all code paths that deserialize SymSpell models, including command-line tools and model loading mechanisms. The issue is resolved in version 3.0.0-M6 by enforcing an upper bound of 10,000,000 on these counts, throwing an IOException if exceeded. Users can raise this limit via a JVM system property if needed.

The vulnerability enables denial of service by exhausting JVM heap memory when loading maliciously crafted SymSpell model files with large count fields. This can crash or halt processes that deserialize these models, impacting availability. There is no indication of code execution or data corruption beyond denial of service. The attack surface includes any component or tool that loads SymSpell models from untrusted or semi-trusted sources.

A fixed version 3.0.0-M6 is available and users should upgrade to this version to mitigate the vulnerability. The fix enforces an upper bound on the count fields to prevent excessive memory allocation. Users unable to upgrade immediately should treat all SymSpell .bin model files as untrusted unless their provenance is verified and avoid loading models from untrusted sources or third-party repositories without integrity checks.