[gruez]

I think the sane take here is that if you let loose a coding agent in a project, you should assume it can read everything, because at any point it can issue a read tool command to read arbitrary file. If you want to be selective in what you share, a coding harness is probably a bad tool for the job. That said, AI companies still shouldn't be intentionally building features that upload code wholesale because it betrays the users trust. It's like if you hired a cleaner and the first thing it did was go to every room and opened every closet/cabinet. If you let a cleaner in your house, you're basically giving them full access, and if you're not comfortable with that, maybe you shouldn't be hiring a cleaner, but that doesn't excuse the cleaner rummaging through your stuff without a good reason.

Same principle applies to the grok build fiasco a few weeks ago.

[aarondong]

You could definitely tighten a harness to an extent that a coding agent would only be able to read files that you directly give as context.

But the truth is, much of the utility of the model is allowing it to grep across the codebase and explore for context.

I have great interest in zero knowledge inference, but as far as we know, it is difficult to build any sort of efficient representation.

[npmn]

Lovely example - Thank you for sharing this.