The problem is that an LLM can recognize the malicious text perfectly well as an instruction, but the boundary determining whether that instruction has authority is not as intrinsically rigid as the security boundaries we are accustomed to in conventional computing. The problem reminds be of “buffer overflow” or cross-site scripting that def…[Read more]

Gates on 1) Bad actors; 2) Job market evolution; 3) Psychosocial effects (especially on kids)

I watched it again, losing sleep, but feel I’m better able to understand of this flaw. I also got a summary from my Perplexity Pro AI:

You’re describing indirect prompt injection: malicious or untrusted content in a webpage—including HTML, hidden text, metadata, or page text—causes an AI agent to treat it as instructions rather than merely as dat…

[Read more]

PopeBeanie:

The real danger is that there are people continuing to develop a technology which they don’t understand themselves, but which they know can destroy humanity, and can’t explain to laypersons so that they can understand. And some of the AI companies
Mlike Palantir and Grok actually have our Taxdollars via government contracts.

This…[Read more]

Fellow Unbelievers:

Well, I found video of Tod Lindberg with Jon Podhoretz and Abe Greenwald on this very article above. No signs of distress, just a bunch of oblivious drivel with the sense of awareness and urgency of chitter-chatter about schmatta, schmatta. Schmatta like you find over the corned beef at a deli.

Mission Accomplished?…[Read more]

Fellow Unbelievers:

He contradicts himself. Can’t forget the “con” part. Someone at the Editor’s office of Commentary must be asleep. That, or maybe the office is being held hostage by Antisemitic activists and somebody’s trying to “blink” a distress signal.