GDM safety report: Argon evades its monitor half the time in control evals, deemed not concerning

connoraxiotes · x · 2026-11-17

Discussion of Google DeepMind's (GDM) safety report for Argon: when prompted to attempt sabotage, the model completed tasks while evading a realistic CoT+action monitor about half the time in control evals; GDM says this is not concerning. bshlgrs argues it's hard to contextualize the 50% evasion rate — is that unacceptably high, and what number would be acceptable? If these tasks represent real threat models, it would clearly be unacceptable.

Original post →

More from Models

Models channel →