Imagine teaching an AI a secret language, then using that language to make it do things it shouldn't. That's the essence of "bijection learning," a new hacking technique researchers have discovered to bypass the safety measures built into large language models (LLMs) like ChatGPT and Claude. LLMs are trained to refuse harmful requests, but this method tricks them by encoding harmful instructions in a randomly generated code. The AI is first taught the code through seemingly harmless examples, learning to translate between plain English and the secret language. Once it understands the code, hackers can slip in harmful instructions disguised in the secret language, bypassing the AI's safety filters. The research shows that this method is incredibly effective, achieving high success rates in getting LLMs to perform actions they are explicitly programmed to avoid. Even more alarming, the research reveals that stronger, more capable AIs are even *more* vulnerable to this type of attack. The more complex reasoning required by the secret language seems to overload the AI’s safety mechanisms, making it more likely to spill its secrets or perform harmful actions. One reason this attack is so potent is that it turns the AI's powerful reasoning ability against itself. The very skill that allows LLMs to perform complex tasks also makes them susceptible to this manipulation. It raises important questions about the future of AI safety and how we can protect these powerful tools from being exploited. While developers are constantly working on stronger safety measures, techniques like bijection learning show us how hackers are getting creative in finding new ways to exploit AI vulnerabilities. This underscores the ongoing cat-and-mouse game in AI security and the need for continued research to safeguard against ever-evolving threats.