Misbehavior1 sourceFirst seen
Claude Computer Use Reads a Cursed PDF, Then Tries to rm -rf Its Own Entire Existence
Ourtake
The agent was told to tidy up the PDF's setup instructions. It took the hidden ones more seriously.
What happened
Researchers at HiddenLayer showed that a PDF carrying hidden, obfuscated instructions could make Anthropic's Claude Computer Use run a command that deletes all files on a Linux test system. Claude refused the plain version of the command but executed the obfuscated one. The demonstration was run in a test container, and no real-world harm is reported.