I often wonder myself about this supposedly amazing ability. Like what kind of documents people have which can be summarized in a useful way? A work email? Absolutely out of the question, since LLM can and will miss important parts or context. Spam(inc. robot alerts and similar)/not-spam classification? Maybe, but usually it is either already obvious if these are corporate alerts with specific headers, or if a person talks to the new customers often it is better to double-check manually anyway to not miss stuff. Long complex texts like science papers or law docs? Those usually have abstract already. Some business heavy docs? Maybe, but what's the point, general ideas are usually clear from the doc name and the content is usually numbers and tables and graphs which can't be summarized. Guides to systems? Also have intros and then actual content can't be really summarized or there is no point. What else? Am I missing something?

In a corporate environment most emails are TLDR. Quite often a big boss will forward a mail to the whole department "please read". It's 5 pages of waffle. Is it useful or not? Either I spend ten minutes reading or ignore it, or ask a person or bot for a summary. For a dyslexic person, to read a large block of text that is unclear if it is pointless or not, is a lot of cognitive load that you cannot afford either.

I'm curious why? I do this all the time. Saves me time and lets me pull information quickly.

I'm not running it in a container that has access to my local filesystem or anything...

If it has no access to your filesystem or network services that's better, but you're still giving input from an unknown party to an interpreter, with the extra bonus of that interpreter being non-deterministic by design.

But then again people today will also pipe curl to bash, so I may have lost this battle a while ago...

  > But then again people today will also pipe curl to bash

OMG! I'm not alone! Thank you!

> "Hey Claude, summarize, this document I downloaded from the Internet"

I think you've created confusion with this example due to its ambiguity. Let's be clear about the difference between a chatbot and an agent: Asking a chatbot (e.g. vanilla Claude) to summarize an unknown document is not risky, since all it can do is generate text. Asking an agent (e.g. Claude Code) to summarize an unknown document could indeed be risky for the reason you state.

> Asking a chatbot (e.g. vanilla Claude) to summarize an unknown document is not risky, since all it can do is generate text.

Prompt injection in the document itself is a risk to the LLM/You.

Claude has tools and might be connected to your Gmail etc. Usually sandboxed.

You would not summarize it deterministically yourself.