CVE-2026-104016

ProtobufParser parses length-prefixed protobuf string and bytes fields by checking ( inputPtr + len) <= inputEnd and then, on the fast path, calling finishShortText(len). Both use 32-bit signed integer arithmetic. When an attacker supplies a varint decoding to a length near Integer.MAX VALUE, inputPtr + len overflows to a negative value, the bounds check passes, and finishShortText(len) reaches TextBuffer.expandCurrentSegment(len), which performs Arrays.copyOf with no cap and throws OutOfMemoryError: Requested array size exceeds VM limit.

The JVM refuses arrays larger than Integer.MAX VALUE - 8 before consulting available heap, so -Xmx tuning does not mitigate the issue. StreamReadConstraints.maxStringLength and maxDocumentLength do not gate this path: a 33-byte payload is far below any document-length cap and the string-length check does not fire before the allocation. This was confirmed empirically against shipping artifacts with maxStringLength, maxDocumentLength and maxNestingDepth all configured.

A February 2026 hardening commit added the equivalent overflow guard to the packed-array-length and message-length paths but left the string-length call sites unguarded. No privileges or user interaction beyond the ability to submit data for parsing are required.