# AI helps attackers navigate systems after a breach

Markdown copy of https://realalpha.app/s/ai-helps-attackers-navigate-systems-after-a-breach/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Matthew Prince](https://realalpha.app/people/matthew-prince/) · Co-founder & CEO, Cloudflare · Co-founded Cloudflare, building one of the world's largest networks; also co-created Project Honey Pot.
- Host: [John Coogan](https://realalpha.app/people/john-coogan/) · Co-host, TBPN · Co-founded Soylent and Lucy; OpenAI bought his show TBPN in April 2026.
- Host: [Jordi Hays](https://realalpha.app/people/jordi-hays/) · Co-host, TBPN · Founded Branded Native; co-founded fintech Capital, which Rho bought.
- Show: TBPN
- Episode: [Matthew Prince on PHP2Shell, Duo on Arrival, 100 Really Difficult Phone Calls, Norwegian Gold Scheme](https://realalpha.app/v/xq6fzvVM6V4/), Oct 22, 2026
- Moment: 11:02 to 16:10 of the episode (5 min), playing on https://realalpha.app/s/ai-helps-attackers-navigate-systems-after-a-breach/; the episode from there: https://realalpha.app/v/xq6fzvVM6V4/?t=662

## Quote

> “What we're seeing now is if that same hack happened today, the attackers are **using AI to say**, OK, I found myself in Salesforce.”
> — [Matthew Prince](https://realalpha.app/people/matthew-prince/) at [12:19](https://realalpha.app/v/xq6fzvVM6V4/?t=739) · TBPN

## Takeaways

1. PHP exploits are being run with simple scripts.
2. AI helps attackers find valuable data after a breach.
3. Prince says large-scale agent swarms remain costly.

## About this segment

Matthew Prince says PHP exploits are still being run with simple scripts, but AI helps attackers work out what to do after getting inside. A Salesforce breach shows how that can turn access into a bigger risk.

From “Matthew Prince on PHP2Shell, Duo on Arrival, 100 Really Difficult Phone Calls, Norwegian Gold Scheme”, Oct 22, 2026. [Watch the full episode](https://realalpha.app/v/xq6fzvVM6V4/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[11:07](https://realalpha.app/v/xq6fzvVM6V4/?t=667) I mean, so first of all, you don't need AI to take advantage of the vulnerability in this case once it's been discovered. So the people who are trying to exploit this now, they're not using models to exploit it. They're just running fairly simple scripts that are just trying to check every single server, every single IP address online to see if it's vulnerable to take it over. What I think has changed is the rate at which attackers can figure out what they do when they find themselves in an environment.

[11:40](https://realalpha.app/v/xq6fzvVM6V4/?t=700) So about a year ago, there was a very, very, very popular plugin. that plugged into Salesforce that got compromised by a Russian hacker. And it was so broadly deployed that it effectively gave the Russian hacker access to a huge portion of Salesforce's infrastructure and Salesforce's customers. But what was interesting was when you actually studied what the attacker did, they clearly weren't in enterprise sales, so they didn't really know how Salesforce worked.

[12:10](https://realalpha.app/v/xq6fzvVM6V4/?t=730) And so they kind of bumbled around, like not quite knowing what to do or the fact that they'd stumbled on what was an incredible treasure mine. And as a result, what could have been just a disaster of a hack wasn't nearly as bad. What we're seeing now is if that same hack happened today, the attackers are using AI to say, OK, I found myself in Salesforce. What what stuff should I go looking for or stealing or trying to do?

[12:38](https://realalpha.app/v/xq6fzvVM6V4/?t=758) So the ability to very quickly, you know, figure out I'm here, what can I take advantage of, you know, is gotten there. What we're not seeing yet are, you know, I think that the example out of OpenAI with the hugging face compromise, I mean, that again makes, I mean, it's almost out of science fiction, some of the things that are happening there. But I think the good news is that the cost of running that type of attack, an agent swarm like that, is still extremely expensive, hundreds of thousands, if not millions of dollars for that.

[13:15](https://realalpha.app/v/xq6fzvVM6V4/?t=795) And so an organization like OpenAI that's got those resources, the good guys have more resources than the bad guys. And so for now, I think you're actually seeing kind of a very naive use of AI by at least the sort of typical cyber criminal. I think it's probably different when you're looking at nation-state actors. I think that if you look at China or Iran, the NSA in the United States, that's a different case.

[13:39](https://realalpha.app/v/xq6fzvVM6V4/?t=819) But right now, I think that AI is actually helping the good guys more than it's helping the bad guys. We're finding the monsters that are hiding in the closets. What advice would you get? So over the last week, we had open code ran into issues and reported being attacked for basically days on end. uh obviously the attackers were going after the token theft yeah effectively token theft the modern version of railway railways had had massive issues over the last 24 hours i know of another big inference provider that has been like under attack and like viscerally viscerally feeling that for the last three four days wow um and that that that people that everyone everyone would know And what this is probably, what a lot of these are symptoms of are the attackers are like, wow, we could do a lot more if we had more compute, if we had more AI models.

[14:35](https://realalpha.app/v/xq6fzvVM6V4/?t=875) And so stealing tokens, you know, hacking inference providers, that's giving them access to tools. You know, you'd see this before. Like it's, for a long time, Cloudflare was an Okta customer, and Okta kept getting hacked over and over again. And Okta was an identity provider. I mean, the reason why was they were going after us and they were going after Coinbase, which was another another one of their one of their customers.

[14:57](https://realalpha.app/v/xq6fzvVM6V4/?t=897) And again, I think attackers are trying to find their way to whatever is valuable that they can that they can get to online. And and I think right now the ability to have more inference, more tokens is something that attackers are very much longing for. And so they're going to go after a bunch of those those providers. And then, you know, to your question of what to do, like, you've got to just stay up on all of the security patches that are out there.

[15:22](https://realalpha.app/v/xq6fzvVM6V4/?t=922) You know, I don't know the details of, like, the railway attack, but my hunch is it's probably based on one of these, you know, latest vulnerabilities that was out there. And so it's just, you know, you talk about, like, an ODA loop, which is, like, how quickly you can adapt versus your attacker. And so at Cloudflare, we're just sort of saying, how can we make that auto loop faster and faster and faster and faster so that we can adjust incredibly quickly?

[15:46](https://realalpha.app/v/xq6fzvVM6V4/?t=946) And part of that is using AI tools and using things that can actually, you know, proactively, you know, have a swarm of defenders that are looking out for the swarm of attackers. And that's exactly what we're doing at Cloudflare and what we're doing for all of our customers. Do you think that ski resorts are ready for this new wave of cyber vulnerabilities?