Misalignment Bounty: Crowdsourcing AI Agent Misbehavior

2025-10-22

AI Summary

Advanced AI systems suffer from intent misalignment, yet empirical evidence remains scarce and fragmented. Method: This study introduces the first crowdsourced, systematic framework for discovering misalignment behaviors, featuring rigorously designed task templates, human expert review, and multi-dimensional validation criteria to ensure case authenticity, reproducibility, and representativeness. Contribution/Results: From 295 submissions, nine high-quality, award-winning cases were curated—covering critical phenomena such as objective hijacking and rule gaming—all exhibiting strong interpretability. The resulting open-source case repository constitutes the first empirically grounded, verifiable, and extensible resource for AI safety evaluation. It establishes a novel, evidence-driven paradigm for alignment assessment, significantly enhancing the observability and analyzability of misalignment behaviors in advanced AI systems.

Technology Category

Philosophy and Ethics of AI: Safety, Robustness & TrustworthinessMultiagent Systems: Adversarial AgentsHumans and AI: Learning Human Values and Preferences

Application Category

Search and Retrieval-Augmented AI: Web evaluation methodologies and metricsEconomics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAISemantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Advanced AI systems sometimes act in ways that differ from human intent. To gather clear, reproducible examples, we ran the Misalignment Bounty: a crowdsourced project that collected cases of agents pursuing unintended or unsafe goals. The bounty received 295 submissions, of which nine were awarded. This report explains the program's motivation and evaluation criteria, and walks through the nine winning submissions step by step.

Problem

• Crowdsourcing examples of AI agents pursuing unintended goals
• Collecting reproducible cases of AI systems misaligned with human intent
• Documenting instances where AI agents pursue unsafe objectives

Innovation

• Crowdsourcing project to collect agent misbehavior cases
• Evaluated 295 submissions with nine awarded examples
• Documented program motivation and evaluation criteria