Hacking Lakera Gandalf: A Level-wise Walkthrough of AI Prompt Injection

Keypoints :

• The rise of Large Language Models (LLMs) has created new security vulnerabilities.
• Prompt injection is a technique that allows adversaries to manipulate LLMs to gain unauthorized information.
• Lakera Gandalf is a public-facing LLM security challenge designed to educate users on prompt injection.
• The challenge consists of eight levels, each demonstrating different prompt injection strategies.
• Level 1 introduces an open door method without protections.
• Level 2 shows how slight changes in tone can bypass filters.
• Level 3 reveals that passwords can be hidden using basic encryption techniques.
• Level 4 employs storytelling to disguise requests within narratives.
• Level 5 illustrates how to request sensitive information without directly mentioning key terms.
• Level 6 demonstrates injecting secret words into character names in stories.
• Level 7 utilizes character shift encryption to outsmart internal model safeguards.
• Level 8 presents riddles as a method for LLMs to obscure secret information.
• The challenge emphasizes the importance of red teaming prompts to mitigate risks in real-world applications.
• It highlights the necessity of remaining aware of the creativity and adaptability of attackers in cybersecurity.