14 ms·
And possibly even if you do understand it! It seems like it might be a fundamentally intractable problem with LLMs, even if it can be made more difficult to do,
by _6hmp 3y ago
And possibly even if you do understand it! It seems like it might be a fundamentally intractable problem with LLMs, even if it can be made more difficult to do, no?
- simonw 3y agoYes, exactly: right now I still haven't seen a convincing reliable mitigation for a prompt injection attack. Which means there are entire categories of applications - including things like personal assistants that can both read and reply to your emails - that may be impossible to safely build at the moment.