Real Alpha
Watch mode

Sabrina Halper Show · 4 min · Oct 7, 2026

AI models still inherit risks from their underlying infrastructure

YC

Yinon Costica

Co-founder, Wiz

Co-founded Wiz, the cloud security company Google bought for $32 billion, and leads its product organization.

  • Sabrina HalperHost · Host, Sabrina Halper Show

    Has worked as an investor at HOF Capital and founded SHS Media.

About this segment

Costica cautions that open-weight models need source scrutiny, then recounts Wiz finding exposed data in DeepSeek’s underlying infrastructure. The example shows AI deployments still inherit cloud and identity risks.

From “Google’s $32B Bet on Cybersecurity in the AI vs. AI Era | Wiz Co-Founder Yinon Costica”, Oct 7, 2026. Watch the full episode

“it's actually on the underlying infrastructure that the model leverages for inputs and training sets”
YCYinon Costicaat 37:22 · Sabrina Halper Show37:22

Takeaways

  1. Model provenance and behavior create supply-chain risks.
  2. A basic scan found an infrastructure vulnerability at DeepSeek.
  3. AI deployments still need cloud, identity, and configuration security.
Watch 3:52 from 34:41

In this episode

Full episode
  1. 8:44Wiz followed customer pain instead of its original idea5:17
  2. 17:47Wiz Atlas makes vulnerability research repeatable5:06
  3. 38:32AI can give defenders a compounding advantage5:00
  4. 47:13Adoption is cybersecurity’s biggest unsolved problem2:40
  5. 44:27Wiz shifted to AI before models were fully capable2:57
  6. 14:09Wiz kept teams aligned by updating its operating framework2:56

  7. 34:41AI models still inherit risks from their underlying infrastructure3:52

  8. 28:38Consumer agents need permission and audit checks3:00Up next

Transcript

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

  1. 34:48

    software on that point i just want to ask you um there is a rise in popularity of like open weight models and this means that anyone can download a model and uh fine tune it and remove the guard rails and use it for any purpose so it makes hackers stronger hypothetically um how do you think like this proliferation of open weight models changes the cyber landscape so it's a great question i think that first there will be a need in the future to create more and more task specific models for for cost reasons for quality reasons for many different reasons for just building owning your own stick in a way i think that on the other side we need to also to be aware that it's a supply chain issue we're getting like a model that we don't know exactly how it was formed we don't know exactly what what it includes and how

  2. 35:48

    it behaves in certain inputs outputs and that's where I think that we if we if people choose to use this we need also to scrutinize the source of it and are we able to actually test for all of these potential impacts in a way that gives us the confidence and the trust that we need now i do think again there is a need that over time will require these open weight models but also we will need to do it in a way that is responsible and attestable at scale so this is one of these things that is still very early and we need to grow the framework alongside the rapid use of adoption and innovation.

  3. 36:39

    Going specifically on DeepSeek, I know you guys found some very interesting vulnerabilities when you looked into it. Can you share a bit about that? Yeah, actually, DeepSeek, I love that example because we tend to think about AI and models an agent, something that is completely detached from the, in the end, from the underlying infrastructure that runs it. And in DeepSeq, the interesting thing is that when DeepSeq went out and we, you know, we wanted to basically understand, you know, basic with this company, how is this product we ran very basic scan and we were able to find this vulnerability that is not in the model or the agent it's actually on the underlying infrastructure that the model leverages for inputs and training sets and we're able to basically access a very large sets of data that

  4. 37:42

    you know should should have not been exposed and they were exposed and i think that this This is a good reminder that in the end, agents and models, they add additional layers of risks, more attack surfaces that now we need to look and secure. But everything in the end runs on infrastructure that still has the same vulnerabilities and misconfigurations and identity related issues that needs to be taken care of and needs to be scanned for and fixed all of the time.

  5. 38:16

    So in a way, we just need to remember that it's a stick. And on top of the cloud and the infrastructure, now you have AI, but you can always target the infrastructure below.