On a related note, an old malicious behavior called an "SQL injection attack" is making a comeback, but it's hyper-charged for the new age of AI. In this attack, a hacker inputs malicious code into an online form in order to gain access to a database or manipulate the #Artificial Intelligence 

Although techies are thinking of ways to thwart this attack, it is difficult to protect against. Probably lots of surface area for this type of attack and the two can used in combination 🤔

https://thegrayarea.tech/the-devious-genius-of-prompt-injection-attacks-on-todays-big-language-models-23ef4ffa6239

The Devious Genius of “Prompt Injection Attacks”

Artificial Intelligence

On a related note, an old malicious behavior called an "SQL injection attack" is making a comeback, but it's hyper-charged for the new age of AI. In this attack, a hacker inputs malicious code into an online form in order to gain access to a database or manipulate the #Artificial Intelligence   Although techies are thinking of ways to thwart this attack, it is difficult to protect against. Probably lots of surface area for this type of attack and the two can used in combination 🤔  https://thegrayarea.tech/the-devious-genius-of-prompt-injection-attacks-on-todays-big-language-models-23ef4ffa6239

Jerome Cody

Re the e-mail assistant tweet, having watched the movie Her I find it a bit scary that people actually want that kind of AI in reality.

Cyril

Interesting, thanks for posting @Jerome Cody 
To me the solution as to how to make any system more robust against these kinds of attacks, is to increase the rigor of the system, by increasing its abilities to reference other information that might be used to cross-check things. This is how we humans work, so no reason it would not similarly work for Ai, I think.

Interesting, thanks for posting @Jerome Cody  To me the solution as to how to make any system more robust against these kinds of attacks, is to increase the rigor of the system, by increasing its abilities to reference other information that might be used to cross-check things. This is how we humans work, so no reason it would not similarly work for Ai, I think.

Sam Mendoza

Exactly!
With these models, code becomes data (and data becomes code)
prompts = code
but is also data
So you have an interface that is Turing complete. It's going to be interesting. Will large corporations, whose CEOs have decided LLM GPT chatbots are the best thing since transistors without having the slightest technical depth, be able to take the right decisions in the next year or so?

Exactly! With these models, code becomes data (and data becomes code) prompts = code but is also data So you have an interface that is Turing complete. It's going to be interesting. Will large corporations, whose CEOs have decided LLM GPT chatbots are the best thing since transistors without having the slightest technical depth, be able to take the right decisions in the next year or so?

Todd Smith

Great food for thought.  I am sure that more fun is on the way with LLMs. “It’s so interesting. I’ve been doing security engineering for decades, and I’m used to security problems that you can fix. But this one you kind of can’t.”

Jerome Cody

A group of likeminded people in Artificial Intelligence are talking about this.