AM-Bench: A theoretically-grounded benchmark for agentic misalignment

Researchers introduced xAM-Bench, a new benchmark designed to evaluate agentic misalignment in AI systems based on theoretical foundations. This tool aims to measure how autonomous agents may deviate from intended goals, supporting safer AI development.