Skip to content

Evaluating Control Protocols for Untrusted AI Agents

arxiv.org
1 pointtimini1 comment
On HN

Comments

This paper evaluates three control strategies for untrusted agents: deferral to trusted models, resampling, and critical action deferral. Initial testing showed resampling and critical action deferral achieving 96% safety. However, adversarial testing revealed resampling crashes to 17% safety when attackers can detect resampling or simulate monitors, while critical action deferral remained robust against all attack strategies.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.