Babel Codes: Large-Scale Measurement of the NUL Ghost Byte and Trust-Boundary Fragmentation in the QR Decoding Ecosystem

Quick Response (QR) codes have become one of the most ubiquitous physical-to-digital bridges in modern computing infrastructure. Their operational security models invariably rely on an implicit and unverified axiom: one code, one message—the assumption that any conformant decoder, regardless of hardware architecture, runtime environment, or software stack, extracts identical payload contents from the same physical barcode. This paper presents an empirical measurement across nine representative decoding implementations, over 50 consumer and enterprise scanning entry points, and 622 systematic payload variants, demonstrating that this foundational axiom systematically fails across the modern computing ecosystem. While the ISO/IEC 18004 specification explicitly permits byte-mode payloads to contain arbitrary 8-bit octets (including 0x00), real-world decoders bifurcate into two mutually exclusive semantic camps when passing decoded results across internal process, language, and framework boundaries: a byte-stream camp (zxing, zbar, jsQR) delivering the intact payload, and a C-string camp (OpenCV Python bindings, Apple Vision, WeChat) silently truncating at the first NUL byte. An exhaustive enumeration of the full C0 control-character space (33 characters, 546 units) proves 0x00 is the unique splitting byte. Combined with the WHATWG URL standard where NUL inside authority is percent-encoded into userinfo, the payload `https://victim.com00@attacker.io/` creates a complete attack loop: truncating security tools validate the legitimate victim domain, while byte-camp client browsers navigate to the attacker. We verify consumer-grade fragmentation on real devices, reproduce end-to-end alert evasion and IOC pollution against TheHive toolchains (ThePhish/ioc-finder), and bypass phishing_catcher scoring. Differential debugging localizes the OpenCV Python root cause to cv2_convert.cpp L.797–800 (PyString_FromString). We also formalize four negative-result theorems (Skeleton-Sharing, RS Sphere Disjointness, Format Info d_min=7, and PUA Mapping Wall), proving that under planar static constraints, NUL is the unique viable ghost byte. Full measurement datasets and replication scripts are released.