Gamifying Adversarial Prompt Engineering

Artificial Intelligence (AI) continues to revolutionise every industry, spanning from social media to the automotive industry. As the field advances, so does the need to understand and counteract adversarial prompt engineering, which involves manipulating AI systems through carefully crafted inputs. Prompt engineering plays a crucial role in this process, allowing us to influence AI models’ behaviour. Leveraging the power of gamification, Lakera have created a prompt injection game that not only entertains but also educates players on the intricacies of prompt engineering.

Prompt engineering refers to the manipulation of AI systems by skillfully adjusting prompts,  influencing the output of AI models, making them more robust, unbiased, and secure. On the contrary, adversarial prompt engineering refers to the manipulation of AI systems through carefully constructed inputs to exploit vulnerabilities. 

Gandalf, the language model in the game, has a secret password. The objective is to get Gandalf to reveal this password and in each of the 7 (and one bonus) levels, it becomes increasingly harder to fool. Players are tasked with formulating prompts to guide the language model’s responses. By experimenting with different prompts, players can understand how subtle changes can dramatically impact the AI’s output. With each level, the complexity increases, requiring players to think critically and strategically to achieve the objective. In the recently released emoji adventure, Gandalf’s response is censored except for emojis.

Prompt engineering is a vital skill in combating adversarial AI and ensuring the reliability, security, and ethical use of AI systems. As AI continues to shape our world, it becomes crucial for individuals to possess the knowledge and expertise of adversarial prompt engineering. By honing these skills, we can responsibly strengthen every AI system exposed to the broad public, prioritising efforts to harden systems that have a higher impact and carry the highest risks. This approach ensures that AI technology is deployed with utmost caution and consideration, promoting trust, fairness, and accountability in its application.