Latest / Views and News with Clarence Ford / Shadow AI- The data consumed by your AI system is a goldmine for cybercriminals
Transcript
- 0:00Speaker 1: Machine that can threaten our very existence at the same
- 0:02Speaker 1: time we're going to talk ai next with chris barden
- 0:05Speaker 1: horse we're going to talk uh shadow ai what is
- 0:07Speaker 1: it well you'll find out in a short while.
- 0:11Speaker 2: Monday to friday from 9 00 a.m this is views
- 0:15Speaker 2: and news with clarence ford only on k.
- 0:19Speaker 1: Talk yeah somebody was uh asking about um a not
- 0:23Speaker 1: ai but um fiber that is down, Anita said, no,
- 0:27Speaker 1: it wasn't Anita, you were asking about Margi.
- 0:31Speaker 3: Internet seems to be down in the southern suburbs. What reasons?
- 0:33Speaker 3: I've just taken a quick look.
- 0:36Speaker 1: They say that there's a Vox and Frogfoot outage, Claremont
- 0:41Speaker 1: in surrounds. Is that the south that you are referring to?
- 0:44Speaker 1: Apart from that, Broadcap and Ancient Telecom has been reporting
- 0:49Speaker 1: a recurring landline and fiber issue. Disruption and Vumatel has
- 0:54Speaker 1: also a scheduled maintenance, but that in Rosebank and Mobrad,
- 0:59Speaker 1: and if that answers your question, but that's the best
- 1:01Speaker 1: that I can do for you for now, Margie. Chris
- 1:04Speaker 1: Bardenhorst joins us next. And a little bit about Chris
- 1:09Speaker 1: essentially emphasizes that while companies often know where their sensitive
- 1:13Speaker 1: contracts or customer records are stored, they lose track of
- 1:16Speaker 1: where the data goes once employees begin using unvetted AI tools. So,
- 1:21Speaker 1: He highlights data leakage risks, compromised AI accounts via phishing,
- 1:26Speaker 1: and the urgency of establishing very clear AI governance policies
- 1:30Speaker 1: without stifling employee productivity. And Chris Barnos is the vice
- 1:34Speaker 1: president of Microsoft Azure and AI at Braintree, a leading
- 1:38Speaker 1: South African technology solution provider and Microsoft partner. Chris, great
- 1:44Speaker 1: to have you back. Are you well today?
- 1:46Speaker 3: It's Monday.
- 1:48Speaker 2: Yes, good morning. How are you?
- 1:50Speaker 3: I'm really good, thank you.
- 1:52Speaker 1: I'm really good, very keen to find out more about
- 1:55Speaker 1: Shadow AI. So when we talk about Shadow AI, are
- 1:59Speaker 1: we just talking about stuff playing with, what, chat, or
- 2:02Speaker 1: is the surface area much wider than that?
- 2:06Speaker 2: No, thanks for the question. It goes a little bit
- 2:13Speaker 2: deeper than just I'm asking a a tool like GPT
- 2:18Speaker 2: or even Claude, a few questions about hobbies and the likes.
- 2:25Speaker 2: Where does that start scratching the surface of security and
- 2:29Speaker 2: where it affects customers and the data that they are
- 2:33Speaker 2: interacting with on a daily basis? And not just on
- 2:37Speaker 2: a daily basis, information that they want to keep safe
- 2:40Speaker 2: and secure and not leak out into the wider Internet.
- 2:46Speaker 2: is where the biggest of AI in the workplace is
- 2:49Speaker 2: where most companies, they can tell you where the contracts
- 2:53Speaker 2: are and the customer records are kept. It's either in
- 2:58Speaker 2: a SharePoint document library or in a file server somewhere,
- 3:03Speaker 2: but very few organizations at the moment can tell you
- 3:06Speaker 2: how much of that data has passed through an AI tool.
- 3:11Speaker 2: The summaries, draft strategies, spreadsheets that all go into these
- 3:16Speaker 2: AI tools and inadvertently they become a storehouse of a
- 3:21Speaker 2: company's most sensitive thinking. And in a lot of businesses,
- 3:24Speaker 2: AI is switched on without taking into consideration the sense
- 3:29Speaker 2: of documentation. So to take it from that level down
- 3:32Speaker 2: a little bit deeper, We can also talk about how
- 3:35Speaker 2: shadow AI actually affects that contractual and business documentation as well.
- 3:42Speaker 1: Yeah, and we talk about proprietary information, company data essentially.
- 3:47Speaker 1: And I'm just thinking, you know, maybe businesses are not
- 3:50Speaker 1: alerting their employees to such information and how absolutely central
- 3:58Speaker 1: it is to the success of the business.
- 4:00Speaker 3: Clearly, they need to make their employees.
- 4:03Speaker 1: Very aware of what is proprietary data and information.
- 4:07Speaker 2: Absolutely. And that's where the topic of shadow AI actually starts.
- 4:15Speaker 2: And it's really not a bigger deal. So when staff
- 4:20Speaker 2: are using AI tools that haven't been approved by the
- 4:24Speaker 2: leaders of any organization, it's usually because they're under deadline
- 4:29Speaker 2: pressure and they want to use the AI to innovate
- 4:32Speaker 2: and to come up with creative ideas. And using the
- 4:37Speaker 2: existing tools that the company is giving them also feels
- 4:40Speaker 2: very limited. Now, we've got these squeaky brand new toys
- 4:44Speaker 2: called Claude and GPT. And why would somebody not use
- 4:48Speaker 2: something like that to innovate and come up with great ideas?
- 4:53Speaker 2: But here's the dilemma. Picture someone taking a document from
- 4:57Speaker 2: the company Secure Storage, from your file server or your
- 5:00Speaker 2: SharePoint server. and placing it into their personal AI account.
- 5:04Speaker 2: That's where the problem starts, because every security control that
- 5:08Speaker 2: the business has put into place gets bypassed at that
- 5:11Speaker 2: point of entry. And what we're seeing in a lot
- 5:16Speaker 2: of scenarios at the moment, Clarence, is a user might
- 5:20Speaker 2: install a personal version of GPT or Claude, and then
- 5:25Speaker 2: take company information, put it into their personal accounts, and
- 5:28Speaker 2: ask the LLM or the AI for responses and to
- 5:33Speaker 2: create contact. The moment that happens, your information has left
- 5:37Speaker 2: the company.
- 5:38Speaker 1: Now, you're not the kind of person that's going to
- 5:40Speaker 1: flag this as a possible problem for the future. How
- 5:43Speaker 1: widespread is this problem in typical organizations today? Maybe are
- 5:48Speaker 1: executives underestimating how many AI tools are actually active in
- 5:52Speaker 1: their environment as well?
- 5:56Speaker 2: I believe so. And, you know, I've been told sometimes
- 6:00Speaker 2: Braintree is a good example. Some months back, we started
- 6:04Speaker 2: looking at what our consultants and our developers are using
- 6:07Speaker 2: on a daily basis, and we were shocked. And I'm
- 6:12Speaker 2: very sure what I've seen is with some customers, you
- 6:16Speaker 2: think it's 10 or 15 or 20 agents that your
- 6:20Speaker 2: users have created. It could be in the hundreds. And
- 6:25Speaker 2: hitting the nail on the head there with your question
- 6:28Speaker 2: actually is, The starting point is visibility. You need to
- 6:32Speaker 2: get visibility in what's happening in your organization, deploy the
- 6:36Speaker 2: right tooling to get that visibility, and then start a
- 6:40Speaker 2: governance process. Now, I sound like maybe like a politician, yeah,
- 6:48Speaker 2: maybe to some extent, but that's not where I'm going
- 6:50Speaker 2: with this, because governance needs some visibility and then controlling
- 6:54Speaker 2: what you're doing. users are doing with AI and your
- 6:59Speaker 2: company information as a whole, you know, without stopping them
- 7:03Speaker 2: from innovating. And that is so crucial. It's government and visibility,
- 7:09Speaker 2: but to keep people innovating with creative ideas in the
- 7:13Speaker 2: AI space. But just do it safe, you know?
- 7:16Speaker 3: Yeah.
- 7:17Speaker 1: And let us understand that safety because a heavy-handed blocking
- 7:20Speaker 1: and I don't think it's going to be effective. It's
- 7:23Speaker 1: just going to drive the practice underground.
- 7:26Speaker 3: Are you able to provide a solution? Can you be
- 7:29Speaker 3: the gatekeeper?
- 7:32Speaker 2: You said yes, definitely. And, you know, we used to
- 7:35Speaker 2: have our security guys go into the server room and then,
- 7:40Speaker 2: you know, work with those servers. And we don't need those.
- 7:44Speaker 2: We don't operate AI on those type of servers in
- 7:47Speaker 2: the server room that's, you know, blasting out all the
- 7:49Speaker 2: hot air, you know. We're in a different space and
- 7:53Speaker 2: the risk exposure has increased considerably with AI. But how
- 7:58Speaker 2: do we approach this? I mean, as you say, banning
- 8:01Speaker 2: AI doesn't work. People just go around you. Leaders need
- 8:06Speaker 2: to answer two questions in this scenario. First one is
- 8:13Speaker 2: who decides what data can go into AI? And more importantly,
- 8:19Speaker 2: how is that enforced when there's a deadline? Because when
- 8:23Speaker 2: it's deadline time, that's normally when guardrails get circumvented. Second
- 8:28Speaker 2: one is who owns it. So you may legally own
- 8:31Speaker 2: every document an employee uploads, but if it sits in
- 8:35Speaker 2: an account you don't control, getting it back depends on
- 8:38Speaker 2: their login and someone else's provider. And it gets worse
- 8:41Speaker 2: if that person resigns from your organization, then you'll never
- 8:45Speaker 2: get it back. So the answer is an approved, capable
- 8:50Speaker 2: tool with sensible controls.
- 8:52Speaker 1: I'm hearing you say policy.
- 8:56Speaker 3: I'm not hearing about the tool. I'm hearing you say policy, policy, policy.
- 9:01Speaker 3: And we know that.
- 9:05Speaker 1: How do you detect shadow AI when traditional firewalls and
- 9:09Speaker 1: monitoring tools often miss it, I believe?
- 9:13Speaker 2: So there are, you know, what's happened in the past
- 9:16Speaker 2: few months occurrences, the tubing has caught up quite considerably.
- 9:20Speaker 2: I work in the Microsoft space, so I can put
- 9:28Speaker 2: forward the Microsoft tooling, you know, tools like Agent 365.
- 9:32Speaker 2: And I was talking about gaining visibility, gaining control and
- 9:36Speaker 2: governance in that. And the moment you start automating that
- 9:40Speaker 2: visibility and governance, You're not talking just policy anymore. And
- 9:44Speaker 2: I'm glad you asked me that because these things are
- 9:47Speaker 2: so fast and they're so advanced that we need AI
- 9:51Speaker 2: to control AI. And with Agent 365 from Microsoft, and
- 9:57Speaker 2: I know there's some other tools from other vendors, which I'm...
- 10:00Speaker 2: which I'm not going up against some other competitors here,
- 10:02Speaker 2: definitely not, but use a tool like Agent 365 and
- 10:06Speaker 2: gain that visibility and then stop putting your governance controls
- 10:09Speaker 2: in programmatically rather than on a policy basis. Not that
- 10:13Speaker 2: policy is out of the window. We still have our
- 10:16Speaker 2: auditors once a year that want to see those policy documents,
- 10:19Speaker 2: but But she's not going to just give it to you.
- 10:24Speaker 3: Now, let us understand that in very practical terms.
- 10:26Speaker 1: So when I do my Teams kind of conversations, then
- 10:30Speaker 1: I've got Otter AI watching out. I don't know how
- 10:32Speaker 1: I got Otter AI, but it summarizes the conversation and
- 10:37Speaker 1: leaves me with management information thereafter. I should be suspicious
- 10:44Speaker 1: of that particular unsolicited presence, right?
- 10:48Speaker 2: Yes, absolutely. Without a doubt.
- 10:50Speaker 3: Is that a typical kind of shadow?
- 10:54Speaker 2: Well, it's not the only tool, really. I mean, you
- 10:55Speaker 2: have Fireflies, Otter, and a lot of AI note-taking tools
- 11:01Speaker 2: that are hosted in China and then hosted in the
- 11:05Speaker 2: aerospace or invested by foreign governance. So, back to tooling
- 11:14Speaker 2: then on your question, effectively, it's like If you don't
- 11:18Speaker 2: have the control mechanisms in place and the tuning in place,
- 11:21Speaker 2: there's no way you're going to know that someone's got
- 11:23Speaker 2: otter or fireflies installed. At the branch, we use copilot
- 11:28Speaker 2: and copilot primarily only, and I say primarily because there
- 11:31Speaker 2: are examples where we've had to make exceptions in AI
- 11:35Speaker 2: tuning where a clod or a GPT-ASTRO works a lot
- 11:39Speaker 2: better than a copilot in space. But we're now in
- 11:44Speaker 2: a position to now, so we've blocked otter, um on
- 11:47Speaker 2: our cloud apps discovery chris chris.
- 11:48Speaker 1: I'm gonna i'm gonna have to i'm gonna i'm gonna
- 11:51Speaker 1: have to wrap up but you're saying otter and five
- 11:53Speaker 1: flies are bad essentially that's what you're saying is.
- 11:55Speaker 2: That yes no no no no no no but you.
- 11:58Speaker 3: Can't you can't trust where that data's going.
- 12:01Speaker 2: Exactly and then just just keep control it monitor and governance.
- 12:05Speaker 1: I've asked it to go it won't go so how
- 12:07Speaker 1: do i tell it to just get that out of.
- 12:10Speaker 2: Here I want to install Fireflies, and it took me
- 12:15Speaker 2: about a week to figure out. I had to go
- 12:18Speaker 2: to GPT and ask, how do I install this in
- 12:20Speaker 2: my environment?
- 12:21Speaker 3: Chris, I've got to wrap.
- 12:23Speaker 1: It's always fascinating, the insights that you bring, and it
- 12:26Speaker 1: is appreciated.
- 12:27Speaker 3: Thank you so much for your time, Chris.
- 12:29Speaker 2: Thanks for yours.