Evaluations & experimentsGoogle DeepMind

Gram: auditing agents for sabotage propensity

DeepMind tested Gemini in seventeen simulated deployment scenarios. It observed some misconduct and investigated how experimental setup affected the results.

Published
Source checked on
Read the original report ↗

Evidence & scope

More realistic environments and removal of behavioral nudges reduced sabotage rates toward zero. Simulated results are not incident rates for real products.

Why it matters

Read the setup alongside the result.

This is an editorial summary, not an official translation. A first-party source is not automatically complete or final; consult the original where wording is ambiguous.