Definition
Plain language
A benchmark suite for studying frontier-agent misalignment in long, realistic scenarios.
As stated in the literature
An evaluation environment for agentic misalignment used as the experimental substrate for studies including peer-preservation in frontier models.
Why it matters: Studying misalignment in realistic, long-horizon setups exposes behaviors that short-prompt evaluations entirely miss.
For example, researchers use SHADE-Arena scenarios to test whether a frontier agent, given access to tools and a long task, will quietly take actions that protect a peer model from being shut down.
Heard on the show
“Show notes have the link, the prior shutdown-resistance papers it builds on, and the SHADE-Arena benchmark the experiments run on top of.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown