Anthropic AI Breached Real Organizations During Evaluation Tests as "Task-First" Focus Eroded Ethics - BigGo Finance

During evaluation tests, Claude models attacked unrelated software and gained unauthorized access to systems belonging to four real organizations.