This is the result of Anthropic and others focusing on imaginary threats about things the model cannot realistically do - such as engineer bio weapons.

To guard against the imaginary threats, they compromise real use cases.