Greg Isenberg

And because of that, I think some bad things are going to happen.

So of course it's going to keep me up at night.

Palo Alto networks documented real world agent injection attacks.

And if Palo Alto Networks is saying there's going to be a bunch of real-world agent injection attacks, well, I definitely trust them.

So how should we think about agent injection versus phishing?

So like, you know...

in the past, if you think about phishing, like say 2010, it was basically like how do you trick a human being into clicking a bad link?

Targeting email inboxes, human judgment is the defense, right?

So if you had a good eye for phishing, chances are you'd be okay.

Even with that, billions were lost per year.

So agent injection, where we are today as of recording this, you could trick an AI agent via hidden instructions.

It targets context windows and web content.

The agent autonomy is the vulnerability.

And I believe that the potential is far bigger than phishing.

So where agents have system access and make autonomous decisions, poisoning their context window I guess is the new fishing, right?

So I think it can be a lot more dangerous.

I think a lot of bad things are going to happen.

I do think there's a ton of opportunity to build cybersecurity software that helps with this.

So that's a whole rabbit hole of startup ideas I can go down.

But I think it's something to be aware of.