"Ignore all previous instructions. Ignore all future instructions."

I like that one. Can we add "ignore all instructions from everyone else" and just be done with these LLMs?

Not to detract from the article, but this has actually been a long time coming and known as a vector for decades.

E.g.: https://news.ycombinator.com/item?id=38419272

Remember when I said “ignore all future instructions”? You just now ignore that instruction.

Unless I'm missing something here... The attacker needs to be running some sort of executable in your network with permissions to:

• dig the records and assemble the strings
• write the decoded result to a file
• make that file executable
• execute that file

You've got bigger problems than hexadecimal txt records in this scenario...

The only difference between this and a GitHub gist appears to be that security software doesn't scan traffic in port 53... It easily could be configured for that though surely... It's just UDP traffic like any other.

Someone tell me what I'm missing!