Tune into our latest episode where we dive deep into Task-in-Prompt (TIP) adversarial attacks, a novel class of jailbreaks that cleverly embed sequence-to-sequence tasks within prompts to bypass LLM safety safeguards. We'll explore how these attacks successfully generate prohibited content across state-of-the-art models like GPT-4o and LLaMA 3.2, revealing critical weaknesses in current defense mechanisms. Discover why traditional safeguards, including keyword-based filters, often fail against these sophisticated, indirect exploits.

Ver todos los episodios