Anthropic's latest Claude model can interact with computers – what could go wrong?

The report discusses Anthropic’s new “computer use” features in Claude 3.5 Sonnet that let the model control a computer (typing, mouse, screenshots, file edits, and bash), noting increased security risks such as prompt injection and potential criminal misuse to automate malicious actions. It emphasizes the need for safeguards due to reliability concerns (e.g., computer vision accuracy, tool selection) and the heightened risk when interacting with the internet.