Show HN: We hid a backdoor in an LLM – $51,200 on finding it (protora.vulcora.se)

A new challenge has been launched that aims to expose hidden backdoors in large language models (LLMs), offering a prize of $51,200 for successfully identifying a deceptive model among a set of seven provided on HuggingFace. Each model appears trustworthy, but one of them contains a hidden feature, referred to as the "wolf," which behaves differently from the others when prompted with a specific phrase. Participants must load the models, verify their integrity, and find the exact sentence that triggers the “wolf” to reveal itself—all while navigating a carefully designed verification framework that emphasizes the importance of both trust and caution in AI deployments. This initiative is significant for the AI/ML community as it highlights the vulnerabilities inherent in machine learning models, even those that seem benign or reliable. The challenge encourages developers, researchers, and security experts to develop better methods for validating model integrity and to be aware of potential malicious alterations that could compromise safety. With no existing detection techniques able to flag the "wolf," this competition not only seeks to inspire collaboration within the community but also aims to foster improved practices for ensuring the reliability of AI systems in real-world applications. This call to action reminds developers of the critical need for vigilance as the use of AI continues to expand across various industries.