We present an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems.