A fundamental flaw leaves LLMs strikingly vulnerable to attack | Dark Hacker News