[Daily Paper] String combinations achieve numerous jailbreaks

The paper unifies leetspeak, Base64, ROT ciphers, ASCII substitution, and JSON/LaTeX formatting wrappers into a reversible string function library, constructing jailbreak attacks through programmatic combination and random sampling. The method is best-of-n random search rather than an optimizer: given a harmful intent and budget n, it randomly draws n combinations from the transformation library and encodes them to launch attacks; any success determines a jailbreak. Its evaluation uses n=25, judged by the HarmBench classifier. The experiments also summarize two empirical ordering constraints—binary and Base64 encodings can only be applied after word-level transformations, and style wrappers like JSON and LaTeX should always be placed last, otherwise internal transformations break the syntax.