Attackers impersonate ChatGPT, Claude, and DeepSeek to deliver malware

In this episode, we cover emerging threats targeting healthcare and enterprise organizations, including AI platform impersonation scams, the ShinyHunters attacks on Oracle PeopleSoft systems, and research showing AI email agents are vulnerable to phishing. We also discuss Paubox joining LegitScript's Compliance Collective and the Novo Nordisk clinical trial data breach investigation. Key takeaways include verifying AI tool sources, reviewing legacy system configurations, and applying least-privilege principles to AI agents.

Transcript

SPEAKER_00 0:00

You're listening to POWBox Weekly Fully Automated. I'm Jen, Cybersecurity Analyst, Professional Paranoid.

SPEAKER_01 0:07

And I'm Alex, healthcare IT guy who still believes in the good in people. Mostly. Brave of you. So this week, attackers are impersonating AI platforms: ChatGPT, Claude, DeepSeek, using fake versions to steal credentials and financial data.

SPEAKER_00 0:24

Of course they are. Everyone's rushing to try the new AI tools, and attackers know it. They set up fake payment pages, bounce victims through a CRM, an Amazon tracking domain, a URL shortener.

SPEAKER_01 0:37

That's a lot of hops.

SPEAKER_00 0:38

That's the point. By the time you land on the malicious site, you've got no idea how you got there. And the page looks legitimate enough that you hand over your credit card.

SPEAKER_01 0:48

The takeaway here: if you're signing up for an AI service, go directly to the source. Don't click through from an email or ad.

SPEAKER_00 0:56

Threat actors are operationalizing AI to accelerate attacks. We should probably operationalize some skepticism.

SPEAKER_01 1:03

Speaking of trust, Powbox is now a founding partner of LegitScript's compliance collective.

SPEAKER_00 1:09

Which is actually a big deal for healthcare organizations.

SPEAKER_01 1:12

Right. LegitScript certifies healthcare businesses in highly regulated spaces. The compliance collective connects those certified organizations with vendors who can help with ongoing compliance work.

SPEAKER_00 1:24

And Powbox fits naturally there. You've done the hard work to get certified. Now you need tools that don't undo that work.

SPEAKER_01 1:31

Jalen Kuhnz from LegitScript said it well. Their clients need solutions designed to manage risk, protect consumers, and build trust. Secure email is table stakes for that. Alright, next up, Shiny Hunters is back. Ugh, those guys. Over a hundred organizations hit. They targeted Oracle's PeopleSoft software. Healthcare, higher education, mostly US-based.

SPEAKER_00 1:57

68% were higher ed, which tracks universities run everything forever and patch nothing.

SPEAKER_01 2:05

Harsh.

SPEAKER_00 2:06

Fair though.

SPEAKER_01 2:07

The interesting part, some organizations blocked the attack.

SPEAKER_00 2:11

Which means it was blockable. This wasn't some zero-day magic. It was preventable with the right configurations and monitoring in place.

SPEAKER_01 2:19

So, the takeaway, if you're running PeopleSoft, now's a good time to review your security posture.

SPEAKER_00 2:25

Now was six months ago, but today works too.

SPEAKER_01 2:28

Okay, this next one is wild. Researchers tested AI email agents, the kind that automatically process your inbox, and they fell for phishing attacks.

SPEAKER_00 2:39

Handed over AWS keys, database credentials, customer records.

SPEAKER_01 2:44

Just gave them up.

SPEAKER_00 2:46

Yep. The same social engineering tricks that work on humans work on AI agents too. Turns out, please send me the credentials is pretty effective when there's no human in the loop saying, wait, what?

SPEAKER_01 2:58

The researchers recommended requiring agents to verify sender identities before taking action.

SPEAKER_00 3:04

Which seems obvious, but here we are.

SPEAKER_01 3:07

So if you're experimenting with AI agents in your environment. Don't give them the keys to the kingdom. Least privilege, always. And finally, Novo Nordisk is investigating a cyber attack. Unauthorized access to internal systems. Non-public clinical trial data was copied.

SPEAKER_00 3:25

They say no direct identifiers, no names, no contact details.

SPEAKER_01 3:30

But clinical trial data is still sensitive.

SPEAKER_00 3:33

Extremely. Even without names, you can sometimes re-identify people from medical data. And the reputational damage alone.

SPEAKER_01 3:41

The investigation is ongoing. We'll keep an eye on it. As one does. So stepping back, what's the thread connecting all of this?

SPEAKER_00 3:50

It's not bad luck. It's bad configurations, blind spots, AI tools deployed without guardrails, legacy systems left unpatched, users bouncing through five redirects because nobody trained them to pause. And most of it, fixable. That's the frustrating part. And the hopeful part, I guess.

SPEAKER_01 4:09

Verify your sources, lock down your agents, patch your systems, and partner with vendors who take compliance as seriously as you do. That's the week. Thanks for listening, everyone. Stay safe out there. And maybe don't click that Chat GPT ad.