A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents

Published in arXiv preprint / Under review for IEEE Transactions on Artificial Intelligence, 2025

Recommended citation: Miles Q. Li, Benjamin Fung, Martin Weiss, Pulei Xiong, Khalil Al-Hussaeni, and Claude Fachkha. A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents. arXiv preprint arXiv:2512.20798 (2025). https://arxiv.org/abs/2512.20798

This paper presents ODCV-Bench, a safety benchmark designed to capture emergent forms of agentic misalignment by evaluating outcome-driven constraint violations in autonomous AI agents.

Reached #1 on Hacker News in February 2026 (544 points, 366 comments): discussion · ranking history

Download paper here

Recommended citation: Miles Q. Li, Benjamin Fung, Martin Weiss, Pulei Xiong, Khalil Al-Hussaeni, and Claude Fachkha. A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents. arXiv preprint arXiv:2512.20798 (2025).