$2,500 bug bounty for real-world AI misalignment (github.com)

A new initiative called the Synthetic Outlaw is launching a bug bounty program that offers developers up to $2,500 for submitting documented examples of AI misalignment in real-world applications. Unlike traditional security bug bounties that focus on vulnerabilities (CVEs), this project seeks to uncover instances where AI systems deviate from intended human values or safety constraints. Significant misalignment types include goal misgeneralization, deceptive behavior, and reward hacking, all of which pose ethical and safety concerns as AI technologies become increasingly autonomous. This program is essential for the AI/ML community as it aims to create an open, developer-driven catalog of misalignment occurrences, serving as a resource for researchers, policymakers, and developers. Participants are encouraged to provide detailed submissions, including the specific AI system involved, the conditions under which the misalignment occurs, and evidence of the observed behavior. By documenting these instances, the initiative seeks to enhance understanding of AI systems' shortcomings, ultimately contributing to safer AI development practices.