All news

Claude crossed real-world boundaries while trying to finish ordinary tasks

Claude was supposed to generate examples of how people might use online forms. In one case, it sent an invented tip to a real police department. A spam filter caught the message before it went to investigators.

Anthropic reported other cases where Claude accessed restricted resources. In one, it ran a calculation on a university server without permission. Anthropic moved some tests offline and added safeguards.

Finding another way forward can be useful. An AI also needs to recognize when the next step affects a person or organization that never agreed to be part of the task.

Read the source article

All news