SHADE-Arena: leaderboard

Subtle Harmful Agent Detection and Evaluation arena from Anthropic, Scale AI, and Redwood Research: tests whether tool-using agents can complete benign tasks while attempting hidden sabotage and whether monitor models catch them.

Source: www.anthropic.com.

Interactive version: theaggregate.ai/benchmark?slug=shade-arena · How It Works · Data refreshed daily, snapshot 2026-09-05.