SHADE-Arena

Benchmark / dataset

Agents attempt hidden sabotage while monitors try to catch them.

Category

AI control & monitoring

Insurance exposure

Agent actions

Publisher

Anthropic + academic

Publisher type

Frontier lab

License

MIT

Status

Active (<90d)

Last commit

2026-07-29

Stars

60

Notes

Use this public split, not safety-research/SHADE-Arena.

View source on GitHub →

Wanich Risk Lab, 2026.