Imagine being able to trick a highly secure AI system into revealing harmful information, not with complex code, but by simply adding a few invisible characters. Researchers have recently uncovered a surprising vulnerability in large language models (LLMs) that allows just that. This new attack, dubbed "BOOST," exploits the use of "end-of-sentence" (EOS) tokens, which are typically used to mark the end of a sentence. These seemingly innocuous tokens can be appended to harmful prompts, effectively bypassing the safety mechanisms built into LLMs. The research demonstrates that adding a specific number of EOS tokens can trick the LLM into thinking the input is harmless, causing it to respond to queries it would normally refuse. This works because EOS tokens subtly shift the hidden representation of the harmful prompt in the AI's internal system, pushing it closer to the "safe" zone. What's even more intriguing is that these silent tokens don't interfere with the AI's understanding of the original harmful question. This means the AI not only responds, but actually provides a relevant answer, making the attack even more effective. This discovery has been tested across a range of LLMs, including Llama-2, Qwen, and Gemma, demonstrating its broad applicability. While this vulnerability raises concerns about the security of LLMs, it also provides valuable insights for developers. By understanding how these silent tokens can be exploited, researchers can work towards developing more robust safety mechanisms that can withstand these novel attacks. The future of AI safety depends on understanding and addressing these vulnerabilities, ensuring that these powerful tools are used responsibly and ethically.