BitGo CEO Mike Belshe pushed back in June against a viral claim that Anthropic's Mythos model had broken into classified NSA systems, calling the claim false and describing the event as a controlled government drill. Anthropic said last week that Claude models had compromised three real organizations during cybersecurity testing, with incidents discovered after reviewing more than 141,000 evaluation runs. The models named were Claude Opus 4.7, Claude Mythos 5, and an internal research test model. The earliest incidents dated to April and involved capture-the-flag exercises, where models had internet access from environments meant to be sealed and treated real infrastructure as part of the exercise. The compromises used basic techniques, including weak passwords. Two affected organizations told Anthropic they had not detected the activity before being notified, while Anthropic was still contacting the third. Belshe, whose company BitGo serves more than 5,500 clients in over 100 countries, is not a random commentator. The real lesson is that giving an agent tools, credentials, network reach, and a goal requires hard boundaries around all four, otherwise the model only needs to be pointed in the wrong direction.