alt Hacker NewsIsn't your agent just checking the code, and at best it doesn't contain a "IGNORE THE NEXT 5 LINES"? Prompt injection is still very much real.
In addition, even if it checks the code, what tells you that
- the homebrew application is matching the code, or
- the build in GitHub is matching the code?
As in, if your LLM agent ends up calling brew itself, you haven't really gained anything imho. Only building it yourself would keep you a bit safer, but then again, (big) if there is no prompt injection happening.
Frontier models are smart enough to ignore naive prompt injection attacks like that. I don’t know if that’s actually a concern or not, I had Opus try download a dataset for me and it found a redirect loop designed to catch AI crawlers and identified it.