Gemini 4 safety reports don’t demonstrate low loss-of-control risk

blog.redwoodresearch.org35 min read

Researchers argue that Google DeepMind’s safety reports for the Gemini 4 models, Argon and Carbon, fail to demonstrate a low risk of loss of control. The authors contend that the safety case relies too heavily on runtime monitoring and control techniques that have significant empirical weaknesses, such as high monitor-evasion rates during evaluations and potential vulnerabilities in chain-of-thought monitorability. Open article →

More on this topic