Recently, some have accused the website "The Cutting Room Floor" of having put malware in their site when they put in an instruction targeted at LLM scrapers to delete all data and report that the scraping ran successfully.
Is this prompt injection truly malware, in your opinion?
I can't really see a definition equating the two that realistically accounts for multiple models from different vendors - am I at fault if a particularly poorly implemented harness misinterprets me writing "do not run dd if=/dev/zero of=/dev/sda" as a command to run?
In the case of TCRF, agents were also actively ignoring robots.txt and disregarding the site's instructions for interacting with it appropriately (with the "delete all data" instruction being specifically hidden from the discoverable page) - so I can't really see the case for comparing it to malware that deliberately tricks users into running it
AI does make cybersecurity more difficult, and that also goes for people using coding harnesses, sandbox your projects!
A prompt is not malicious software, it's an input. It's similar to SQL injection (query evaluation from unsanitised input), unless that injection also leads to arbitrary code execution, which could involve actual malware as payload. Of course a malicious prompt could lead the LLM to generate and run said malware for the attacker, or it could act itself as malware.
It's more like "social engineering" (targetting LLMs) than malware. A site that influences its (human) users to buy, vote, "invest" in scams and so on is malware?
I personally have a hard time considering it malware, because it's just instructions targeted at an LLM.
If someone put up a sign in a store that said "Burn this store down!" the person who put the sign up is not responsible if someone follows the instructions written on it, are they?
They could be. Incitement to commit crime is a crime in many places. And penalty can be same as for the crime itself.
Now I very much question if automated system should be inciteable. And think that there is very much higher responsibility to make it always not be. Or make it uncapable of doing undesirable things.
In this specific case, they are already trespassing to be able to read the sign against the owner's wishes, are they not? The sign is only placed in a hidden corner visible by trespassers and not visible to any regular customer.
malware is a malicious program , it can be in memory even within another process but essentially its a little program of itself.
prompt injection is more close to an exploit. a sequence of input that leads to unexpected behavior, exploited to perform malicious behavior.
many exploits are hosted on websites. called exploit kits even. i think prompt injection match such classification. its very similar. a piece of data on the site's side exploits a program or interface on client side.
In the case of TCRF, agents were also actively ignoring robots.txt and disregarding the site's instructions for interacting with it appropriately (with the "delete all data" instruction being specifically hidden from the discoverable page) - so I can't really see the case for comparing it to malware that deliberately tricks users into running it
AI does make cybersecurity more difficult, and that also goes for people using coding harnesses, sandbox your projects!
If someone put up a sign in a store that said "Burn this store down!" the person who put the sign up is not responsible if someone follows the instructions written on it, are they?
Now I very much question if automated system should be inciteable. And think that there is very much higher responsibility to make it always not be. Or make it uncapable of doing undesirable things.
prompt injection is more close to an exploit. a sequence of input that leads to unexpected behavior, exploited to perform malicious behavior.
many exploits are hosted on websites. called exploit kits even. i think prompt injection match such classification. its very similar. a piece of data on the site's side exploits a program or interface on client side.