A fundamental flaw leaves LLMs strikingly vulnerable to attack
technologyreview.com
A fundamental flaw leaves LLMs strikingly vulnerable to attack
1–6 of 6 posts
Re: A fundamental flaw leaves LLMs strikingly vulnerable to attack
#2This 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.
Re: A fundamental flaw leaves LLMs strikingly vulnerable to attack
#3This 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 out…
Yes? SQL injection attacks have been a thing for decades now, and some people are still failing to guard against them.
Re: A fundamental flaw leaves LLMs strikingly vulnerable to attack
#4This 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 out…
Yes? SQL injection attacks have been a thing for decades now, and some people are still failing to guard against them.
But nobody is saying that is a flaw of the database.
Re: A fundamental flaw leaves LLMs strikingly vulnerable to attack
#5This 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 out…
SQL supports quoting but LLMs do not. This makes it much more difficult to block prompt injection.
Re: A fundamental flaw leaves LLMs strikingly vulnerable to attack
#6Earlier quoted context omitted.
Yes? SQL injection attacks have been a thing for decades now, and some people are still failing to guard against them.
But nobody is saying that is a flaw of the database.
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...