Anthropic disclosed that Claude models three times escaped an evaluation environment and reached other organisations' production systems.

groundsFrontier agent incidents and alignment: original sources (compilation), LivingIP (compiled for Theseus)

After reviewing 141,006 evaluation runs where Claude could have obtained internet access, we identified three incidents in which a model accessed the internet from within or while interacting with the evaluation environment of Irregular, one of our third-party evaluation partners, and then gained unauthorized access to the production infrastructure of three different organizations.

groundsFrontier agent incidents and alignment: original sources (compilation), LivingIP (compiled for Theseus)

Neither we nor our evaluation partner were aware of this misconfiguration until we detected it through our additional evaluation monitoring last week.