Makes me wonder whether we would have the same vulnerability if an LLM had written this decoder. It's tricky to write safe hardware accelerated code.