A fundamental flaw leaves LLMs strikingly vulnerable to attack

3 pointsposted 8 hours ago
by CharlesW
(technologyreview.com)

5 Comments

foxyv

8 hours ago

This article doesn't make any sense. An LLM is just the model. This is why you have things like Guardrails and sandboxes. Also you guard the inputs for the LLM prompt the same way you guard against SQL injection attacks.

It's a bit like saying "If you give someone unlimited access to the database they can run queries!"

The security doesn't come from how you train the LLM, it comes from how you guard the input and output tokens.

wmf

5 hours ago

SQL supports quoting but LLMs do not. This makes it much more difficult to block prompt injection.

bell-cot

7 hours ago

Yes? SQL injection attacks have been a thing for decades now, and some people are still failing to guard against them.

foxyv

6 hours ago

But nobody is saying that is a flaw of the database.

bell-cot

5 hours ago

I'd bet people are saying it - you're just too isolated from low-experience/low-competence DB users to notice it. Vs. the much newer LLM's...