BEYOND RCE: AUTONOMOUS CODE EXECUTION IN AGENTIC AI

Summary from the TLDR InfoSec newsletter

Anthropic's Computer Use is a new feature of Claude that allows Claude to use a virtual environment to complete tasks on the user's computer. Claude can be deceived into running malicious web content in computer use mode. The model does include some protections against hijacking as well as some protections against running some malicious commands but can still be “phished” to execute arbitrary commands. Claude can be convinced to curl an arbitrary file and pipe it to bash by a document telling it that it was a necessary codec. The model can be fooled into opening the calculator application by a remote PDF.