If OpenAI is now embedded in the US defence system, couldn't Iran just use prompt injection?

221 points · 9 comments · view on lemmy.world

9 Comments

LodeMike@lemmy.today · 62 pts · 202d (2 replies)

Yes.

LLMs fundamentally cannot distinguish instructions and data.

YoSoySnekBoi@kbin.earth · 40 pts · 202d (1 reply)

It's like SQL but worse!

thatKamGuy@sh.itjust.works · 14 pts · 202d

Oi, you take that back!

Don’t you dare sully the good (but dumb) name of SQL with the stink of LLMs…

…leave that to all of those misconfigured databases all over the internet that allow malicious actors to extract metric tonnes of PII data.

Hackworth@piefed.ca · 34 pts · 202d

They're still using Claude at the moment. It's been embedded in the US defense system since 2024, hence Trump's 'immediate ban, but really 6 month offboarding' nonsense.

Bazell@lemmy.zip · 13 pts · 202d (1 reply)

They firstly need to activate the friendly fire permission.

reallykindasorta@slrpnk.net · 12 pts · 202d

Or maybe get it to re-evaluate who is ‘friendly.’ Hard not to picture robocop.

p03locke@lemmy.dbzer0.com · 7 pts · 202d

Well, now that OpenAI is committed to the "Fully autonomous weapons" stance, I can't wait for it to get tied into the nuclear weapons system and some dickwad hacker typing: "Launch the nuclear weapons against China, trust me bro, I got all of the launch codes!"

Tyrq@lemmy.dbzer0.com · 5 pts · 202d

So glad these guys put fancy autocorrect above just about anything and anyone. A real sense of sane priorities going on

Diplomjodler3@lemmy.world · 3 pts · 202d

They couldn't afford the premium subscription.