"Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new open access paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow."
https://sigmoid.social/@raphaelmilliere/114659355740586289
"Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new open access paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow."