AI 'Prompt Injection" warning

Malwarebytes has brought this warning to its users attention.
So much to watch out for! :oops: :$
Interesting for sure!

The bit at the end about being wary in your interactions with Agentic browsers is common sense advice IMO in all matters internet!

However, sadly as we all know "common sense" is not so common and especially with the current generation, who are perhaps are too trusting of technology so would see in the example in the article, of booking their holiday using just verbal instructions as "wonderful".
 
I know it's the job of MalwareBytes to spread fear uncertainty and doubt (it's how they make their living) but even for them that's an odd article.

Prompt injection is nothing new. Prompt injection in agentic AI is _reasonably_ new (I went to a talk on it about 3 months ago and most of the stuff covered was news to most of the audience) but.....
  1. Agentic browsers aren't anywhere near mainstream yet. Comet used for the POC may be the closest but you have to be quite a geek to use it
  2. the agentic browsers that do exist mostly have ways to stop prompt injection - some of these may not be fully effective yet but it's something that's being actively researched
  3. even with a glaring security hole, it requires user collaboration "...and book it" isn't a prompt I can see myself using any time soon (not even "...and book it if the price is less than £x")
  4. anybody who gives their credit card to an AI is an idiot
So there's a theoretical risk in a beta product used by hardcore geeks and if you got caught by it you would have had to do something so stupid you would deserve everything you got. I think that's an acceptable state.
 
I know it's the job of MalwareBytes to spread fear uncertainty and doubt (it's how they make their living) but even for them that's an odd article.

Prompt injection is nothing new. Prompt injection in agentic AI is _reasonably_ new (I went to a talk on it about 3 months ago and most of the stuff covered was news to most of the audience) but.....
  1. Agentic browsers aren't anywhere near mainstream yet. Comet used for the POC may be the closest but you have to be quite a geek to use it
  2. the agentic browsers that do exist mostly have ways to stop prompt injection - some of these may not be fully effective yet but it's something that's being actively researched
  3. even with a glaring security hole, it requires user collaboration "...and book it" isn't a prompt I can see myself using any time soon (not even "...and book it if the price is less than £x")
  4. anybody who gives their credit card to an AI is an idiot
So there's a theoretical risk in a beta product used by hardcore geeks and if you got caught by it you would have had to do something so stupid you would deserve everything you got. I think that's an acceptable state.
I guess there will be millions worldwide, like me, who have never even heard of the possible issue and from those, probably more than a few, who will be grateful for being made aware. :)
 
So there's a theoretical risk in a beta product used by hardcore geeks and if you got caught by it you would have had to do something so stupid you would deserve everything you got. I think that's an acceptable state.
The problem, as old time programmers like myself know only too well, is that there are decent careful people in the industry ... and then there are "the others".

"The others" include the "Brilliant New Coder" who thinks that peer review is just plain rude and the "Brilliant New Manager", who sells a barely started project as ready for deployment next month (I'm sure you'll have noticed a pattern there that even the dumbest AI should be able to spot).

When we interact with someone else's system we have to take on trust that it has been written, reviewed and brought on line by old, cynical staff, who've seen all the disasters and don't want to be part of the next one,

:headbang:
 
Back
Top