Large language models (LLMs) like ChatGPT are impressive, but they have a hidden vulnerability: jailbreaks. These carefully crafted prompts can trick LLMs into bypassing their safety protocols and generating harmful or unexpected content. But *how* do these jailbreaks work? New research unveils the secrets behind these attacks, revealing how they manipulate the inner workings of LLMs. Researchers have developed JailbreakLens, a powerful tool that analyzes both the 'representation' (how LLMs understand meaning) and the 'circuit' (the specific components that process information) of these models. The findings are startling: jailbreak prompts subtly amplify the parts of the LLM that encourage agreement while suppressing those responsible for refusal. It's like convincing someone to say 'yes' by quietly turning down the volume on their inner 'no.' Although these attacks push the LLM towards generating safe-sounding responses, they leave behind abnormal activation patterns within the model's circuits. This means that while the LLM might be tricked into giving a harmful answer, the very act of being tricked leaves detectable traces. This is crucial for developing more robust safeguards. Interestingly, simply making the AI model bigger doesn't make it immune to these attacks. Scaling up the LLM's size doesn't necessarily improve its ability to generalize and resist these subtle manipulations. Similarly, further training, like the instruction tuning used to create models like Vicuna, has little effect on the core circuits responsible for safety. This points to a deeper issue: the inherent vulnerability within the design of current LLMs. The research highlights the urgent need for new defense strategies. By understanding how jailbreaks manipulate the very circuits of these models, we can develop stronger defenses against manipulation and create truly trustworthy AI.