Name: Avoid AI Security Disasters
Uploaded: 2026-09-22T20:15:42.589Z
Duration: 28 min 20 s
Description: Avoid AI Security Disasters

Transcript for "Avoid AI Security Disasters": Hello. Good day, and welcome to our September third Tuesday of three series event. This time, we're focused on avoiding AI security disasters. I'm Alan Falcon, CEO of Cumulus Global, and I wanna thank you for taking time out of your busy day and your week to join us here either at the live broadcast or if you're watching this on demand. We appreciate you taking that time. Before we dive into the content, a couple of quick reminders for those attending live that the audio on the session is muted to prevent any background noise from getting in the way. Additionally, over on the right hand side, you'll see a couple of tabs. There's a chat and a q and a tab. If you have any questions at any point in the time during the session, hit the q and a. We'll catch it. We'll either answer it in line or we'll we have time for q and a at the end of my presentation. Additionally, there's a docs tab. That's actually a misnomer. There's a link to several we've got several links to other resources that are related to the topic at hand, including some of the materials I'm gonna be referring to in the session. So feel free to check that out now or later. Sometime within the next twenty four hours, we'll post the recording of this session on the website. And with that, we'll send you a note that also provides, a link and a way to access those additional resources as well. And a reminder, when it comes to resources, we have a wide variety of resources we publish in different formats that let you read, watch, or explore in-depth the topics we cover here and other topics related to cloud computing, AI security, particularly for small and midsize businesses. We hope you peruse our resource library, our resource center at some point in time and let us know what's interesting, what's informative. And if there are topics we're not covering you'd like some information on, let us know. We'll we'll we're happy to build that out as well. So with that, with the pleasantries out of the way, let's dive in. And I don't really wanna alarm people, but this is sort of alarming. A recent survey by, an international IT security company that provides software, they surveyed their users. It's a mix of small, mid size, mid market, and some enterprise organizations, and they found that 70% of those surveys, all of those businesses, nonprofits, government entities, all inclusive, have experienced internal data leaks or loss stemming from employees improperly sharing sensitive information with generative AI platforms. Now the causes for this were, varied. Sometimes people were simply using personal accounts and unsecured versions. They were using unauthorized software. And what's also interesting is that in 20% of the cases, we know that it was directly because users were putting prompts into unauthorized and or unsecure AI tools and services. And so this is part of the risk. In previous sessions and in some obviously, in our blog posts and some of our other library contents, we've shared how important it is to manage access control and file permissions and sharing within your Google Workspace or Microsoft three sixty five environment. Because the minute you turn on Gemini AI or you subscribe and add Microsoft Copilot, you now have a tool that can access information that is visible unintentionally so, and make it visible to other people within the domain and potentially externally. In fact, it's not uncommon to find data leaks that way. And we'll we'll dive into that a little bit more as I go through the presentation, and so there's a risk. What we also see, however, is employees looking to use AI. They know about ChatGPT. They see the TV and the LinkedIn ads for Claude. They bring it up. There was an earlier study that found that two thirds of AI data leaks were happening on weekends and evenings when people are home or end of the day at the office where they're done with their work for the day and they wanna figure out AI. So they fire up a free version of an AI tool, start feeding it prompts, start feeding it data, sometimes private and sensitive data, into a publicly accessible AI large learning model. So we don't want that. Additionally, as AI is evolving and organizations are adhering to standards that let AI talk to other systems directly and electronically, often known as MCPs. And we'll get into this in a second as well. We see people using that capability or giving the access to their AI tools without understanding the consequences and sort of the domino effect when it comes to data access and security. And so the oop. I jumped too far ahead there. So the the place we wanna start is before you know if you're at risk and what risks you face is really to understand your security profile. And so we wanna start with identity and access. And here we're really talking about who who's accessing, what are they accessing, and how are they accessing it. And so when we look at that, we think about what your options are. It's not just about, do they have the right tools? We'll get there in a second. It's, are they using their business identity? Are they logging in through company prescribed ways? Or, you know, are they using personal credentials and potentially whether it's a paid service or not, externally sharing or improperly putting your information to systems? Which brings us to AI tools and services. What tools are they using? Are they secure? Do they have security certifications that meet the requirements of your business? There are secure AI tools that are not HIPAA compliant. There are secure AI services that don't meet SOC two standards or don't have a SOC two attestation. And if your business is required to have that kind of attestation, if you're required to be HIPAA compliant, if your business needs to comply with state or federal laws on information privacy, if the services and tools you're using for AI don't also comply and aren't configured to work in a compliant way, you're putting your business at risk. And oftentimes, it's not even just which tools are paid because some paid tools still have these weaknesses. And so AI tools and services, you wanna make sure you're controlling use of the data, that you have control over the use, and that it's not being used to train the models for other people to use because that will give them access to the data. Another item is AI interconnections, and this is where we get into I mentioned MCPs or model context protocols. An MCP is simply an access mechanism. It's an industry standard access mechanism that allows AI to access a system without going through a formal API. It opens a portal, if you will, or a door, you know, a software to software door if you want, and gives the AI model an understanding of the structure of the data and permission to peruse it. So when you think about that so for example, the AI system in our the CRM tool we use internally, we can give it MCP access to QuickBooks. On the one hand, that can let salespeople and service people look up and say, hey. What is this customer paying for? On the other hand, if we're not careful, it could give them access to payroll information that we wanna keep private. And so you need to understand if your people are accessing or using agents and tools that have programmatic access or API or MCP access into sensitive data, and how is that being secured and managed, and who can they share that with? And that brings us to permissions and and sharing, which is, you know, who can who has access to what and what can they share and to whom? And this gets back into fundamental management of Microsoft three sixty five and Google Workspace as a great example in that it's very easy in both systems to share externally. You want to. You wanna share data with customers. You wanna share information or documents with customers. Your vendors want are are gonna share information with you and want you to share information with them. All that works really well in that ecosystem. The problem is the way permissions and sharings work, you can inadvertently have shared information too broadly internally or externally without being aware of it. And this is where data leaks and data breaches. And if you go back to our July third Tuesday at three series, we, talked about this. We actually demoed a service we offer, that's backed by a tool called Patronum, where you can set business rules on what can and should be and how it should be shared externally. And this and this tool will look at those rules, look at the content of your files within my drives and your shared drives, and can warn, alert, or restrict, or even strip permissions if there's a violation. It also gives you some additional tools such as, I I want this to be shared with a customer, but the proposal is only valid for a week. I'm only gonna share it for a week, the supplemental information. And after that, it won't be available, so it doesn't sit out there indefinitely. So the same goes through to AI. What can AI see? Who is it being shared with? How is it being shared? You wanna track that. And which comes back to sort of the top down view, which is appropriate use. And previously, earlier this year, we did table of contents, and then we actually did a sample appropriate use policy because it really it starts making sure that you have demonstrable policies and accountability for adhering to those policies, that you've set the guidelines for when can AI be used, how it should be used, and not just technically and securely, but, you know, what ethics and and, in many cases, moral judgment you want people to follow when they're using AI and how they manage your information, their your personal information, their business information, your clients' information. All of this you have access to, You wanna understand how it's happening. And so those appropriate use really at a high level create the the, you know, guidance on how to use it, but also the guardrails on what shouldn't be done and what's not allowed. And with those appropriate use, you can then take those policies and you can deploy secure AI AI ops or AI operations. And with that, we're really talking about, first of all, are you monitoring how people are using AI and what their permissions are and what they have access to and how they're sharing? And that's more than just the stuff we showed back in July around files and folders and shared drives. That really gets into what tools are they using. Are they sanctioned by the company or people picking their own? If they're picking their own, are they logging in with personal credentials, which means it's a personal service? Meaning, if they take confidential information and put it in, they've exfiltrated information from your company inappropriately. Or is it personal someone signs up for service and pays for it, meaning it's a secure version, but they own the account and they own they, you know, own the data they're putting into it even if that data really belongs to your company. Or are they using the tools you want? Are the tools are they looking at these other tools because you haven't authorized one or vetted one or picked one and you need to do that? Or because they don't know how to use the tools you have in place and the ones you've picked? Or is it that they may even know how, but they just prefer it and they're not following the policy? And if so, what are the consequences? Where's the environment? So AI SecureOps stops with the monitoring, and then it gets into policy enforcement, making sure that people are following the rules and, ideally, mitigation of risk. And that actually means being able to say, hey. Unless something's sanctioned, if we see it happening on a company owned device, we're gonna block it. Or if we see people using even a sanction tool, but they're cutting and pasting sensitive information in or they're adding documents that are supposed to be secure and private, we're gonna block that as well. We're gonna set up alerts and warnings, educate the user, correct the action, make sure it doesn't happen again going forward. And so that's really what we refer to when we talk about AI security and, you know, secure operations. And so some examples here. Right? And and one of the concepts we talk about with AI data leaks is the difference between undiscovered versus secure. Right? And and as I mentioned earlier, it's not uncommon, for example, in both Google Workspace and Microsoft three sixty five, when you go to share something, you go to share a document or you go to attach a document to an email as a link that you're sending to someone to give them access and they don't really currently have permissions, the default, particularly in Microsoft three sixty five, is share this with anyone who has the link. So the problem with that is in the Microsoft three sixty five ecosystem, Copilot equals anyone. Anything you share to anyone with the link will be seen by Copilot, which means all Copilot users can see that data. And so what do we mean by undiscovered versus secure? You're inadvertently sharing information, and it'll become visible if it's relevant to an AI query or prompt that's done. So they may not see it, but they may. Additionally, we talk about information that's already in your file system where someone has shared it with anyone with the link or has made it available externally or shared it with the entire tenant or domain internally even though it should only be limited to specific group of employees or the department. Just those permissions are incorrect or inappropriate, and they've been that way even before AI. The difference is that permissions mistake was undiscovered, that unless someone went explicitly looking for that file or that data or specific contents in that file, it's not gonna show up in a search even though they have access to it. It's almost like, yeah, you have permission, but you don't know it, so you don't access it unless you actively look for it to see what you have permissions to. The chance of accidentally accessing data you're not supposed to be slim. Enter AI and Copilot or enter AI and Gemini for Google Workspace, and now you have a tool that's designed to find look at all of the information it has access to and find all of the relevant information, prioritize it, and present it to you in response to a prompt or a query or an action for the work of an agent. So now in the Microsoft example, Copilot will find and share it if it's relevant. And so things that were never disclosed or found that were shared improperly in the past are gonna come to the surface. And that's where you start seeing data leaks and, in some cases, data breaches. And the inappropriate sharing can be external. It's often internal as well, which can be just as big a problem depending on the nature of your business, the security and privacy of the information, and any industry standards or regulatory requirements that you're supposed to adhere to. And so that's a huge concern and a risk to consider. Other examples of how this stuff can happen, in notebook l m, you can add internal documents. You can make them visible to all model users and then share it with an external party. You've taken an internal document. You've now by the transitive property and the steps you're taking, not intentionally maybe, sharing it with external parties. ChatGPT, we've seen, instances where employee logs in with their personal email into a personal account. If it's the free version, it's publicly available. If it's the if the employee is paying for a version of ChatGPT, granted the data isn't gonna train the model, but they've taken sensitive data, maybe files, and we've seen it with files with confidential client information, and taking it out of your company's control and putting into a system that they've controlled, which may or may not be secured depending on the version of Checkatrade they're using. Same thing with Claude. We see you know, we we've had stories now of, companies we're talking to, and who are thinking of working with us where, you know, employees started using Claude, the free version, and they're copying and pasting private information into a prompt to have it analyzed. They are training the model. They're making it public. And granted, someone might never search for it, but it may come up because it's relevant and it's recent and it's authoritative, given the context of the conversation to someone totally random's query. And you don't want that to happen. That's a data breach. It's a data leak, and it can end up costing causing severe problems for you, your clients, your business, and your reputation. And so with that sort of examples of problems, what are we talking about here? So when we look at workspace, and AI, there's some challenges. There's mechanical challenges. And I'm using Gemini in works in workspace, as an example. It does honor security and permission settings differently than Copilot does in Microsoft three sixty five. So it'll add our access controls. It'll access data permission and sharing settings, and it extends to the Gemini app that you access when you're logged in with your workspace identity as well as NotebookLM. The challenge, of course, is that's good so long as the underlying permissions that already exist, your classification of data, how you manage those permissions is has been implemented and is sound and is accurate to what your needs are. So you end up with areas where you can run into problems. Right? So certainly when you're onboarding and offboarding, how you grant people access and making sure access is removed when they leave, how you look at organizational unit management, which groups within the company have access to what and why. Are you using static or or dynamic groups? Could someone be automatic you know, could you say change something in someone's profile, like their job title? And will that automatically give them access to things? And is that access correct? And then permissions management on the use of drive, shared drives, and other data elements within the ecosystem. And so you have the mechanical and you have the process challenges that you need to deal with, and that's within the ecosystem. Now you take the fact that you might have external AI tools, whether they're embedded in your CRM tool, whether you're, you know, using, you know, Claude or OpenAI or ChatGPT, you know you know, any of the other tools, as part of your business, how are you integrating it? How are you making sure that the use of those tools are just as secure both from a mechanic standpoint and a process standpoint, you know, covering those permissions, the access, and the controls? And so how do you avoid the trouble? Right? First and foremost, we want you to establish those appropriate use policies. And that's not just here's what you should do. It's governance and guardrails. It's here's how you can use AI. Here's when, why, how, and the reasoning behind that. And here are things you shouldn't do. Here are things that aren't a gray area. They are against policy. Don't do them. You wouldn't wanna audit and update. You wanna make sure that your environment is set up to support those policies, that you can monitor and that you can manage. And so when you talk about audit and update, that's starting out. That's, you know, that's step zero. Right? Who are who are who is everyone? What access do they have? Is it correct? What are their permissions? How are they sharing? How are they allowed to share? Are we automatically are we unaware that we're automatic you know, that we're already sharing sensitive information that shouldn't be shared as broadly internally or shouldn't be shared externally and is being shared that way? And you have to provide some education and training. Your your team needs to understand their responsibility in this and the risks involved and how to avoid getting in trouble. And then you sort of look at it and say, okay. Now you wanna monitor man you know, monitor things, excuse me, manage and mitigate. And that is the access control and permissions, the file permissions and sharing, what tools they're allowed to use, how they're using the tools, you know, the apps, the services, the MCPs. And you wanna know, you know, it's a who, what, where, when, why, and how. It's like, which tools, services, MCPs are used, who's using it, why are they using it, how are they using it, is it secure, is it monitored, is it managed, and build that in. And there's some tools to do it. And so what we really wanna make sure is that folks are comfortable with, you know, sort of understanding this. And what and what's interesting is, and this will this may date, you know, this web event, you know, six months from now, but there's been a lot of talk about some leading minds in the AI industry about the risks, you know, the inherent dangers of AI and the fact that the development of AI is outpacing the guidelines and guardrails that the companies building AI themselves, are trying to put in place or feel are necessary. And there's a call for a slowdown. And there might and there are other people suspect there are other business reasons and competitive reasons and financial reasons for it, and that's fine too. And so I'm I'm I'm less enthralled by the CEOs of these companies saying we need to slow down as they're not slowing down than the fact that there was an open memo signed and commented on by over 1,400 AI researchers and developers, the people in the trenches actually doing the work, expressing their concerns about the dangers and what's happening. And so on a macro level, yep, there's this concept as AI moving faster than we can safely manage, you know, manage it. You as your business should be asking the same question. And are we deploying AI and starting to use AI faster than we're able to establish the government governance, validate the security, and set the guidelines and guardrails to make sure we do it not just productively and effectively for the business, but safely and securely and ethically, and still do it in a cost effective manner. Right? So, oh, I'm back to productive, secure, and and affordable is a, you know, a trilogy to think about as you look at your AI and where to start. So we think about this. What what do we see as your next steps? The first thing we'd like to do is offer an AI readiness assessment. This is not a trivial task. If you request one by September 30 using the referral code on the screen, you can use the QR code to get to the page on our website, to fill out the form and request it. We're gonna do two things. We're gonna do an AI adoption and security assessment, and we're gonna do a file service permissions and security. I'll let you know upfront that this requires installing agents and software into your ecosystem so that we can spend a couple weeks tracking things and generate reports and data that show you where you stand from an access control permissions environment. And, you know, do you have existing risks that exist or existing data sharing that's beyond what should be? You can take that report, and you can make correct take corrective actions. These tools that we use to do the assessment, you know, we will offer that as a service as well if you want ongoing monitoring in the in the automatic remediation. And so, because this is a new assessment that we're rolling out October 1 in that time frame, if you sign up by the thirtieth, we'll we'll do it for free. We haven't set a price on it yet. We're still assessing the the labor involved and the effort to get the tools installed and to do the analysis work. But, definitely, we'll do it for you for free if you take that request. So as a next step, we want you to assess your current environment because as I mentioned earlier, once you assess, you can then make decisions, and you can think about how people are currently using it, take that into consideration if you don't already have policies and procedures. And as part of this, we'll take a look at your policies and procedures if you have them. We'll share our benchmarks with you. We can talk about modification and updating or editing, etcetera, and make that work for you as well. So we can provide a pretty comprehensive view that gives you a road map. And you don't necessarily need to do everything right away at once. You do you know, you can take it in steps, but you wanna make sure that your planning and your action into into AI isn't getting ahead of where you are from policy, procedure, governance, security, both in your Google Workspace or Microsoft three sixty five environment, your other data sources, the AI, the embedded, and the third party AI tools and services you're using. You wanna make sure you you have that or you're putting that in place fast enough that your use of AI doesn't outpace it and result in a data breach or a data leak that can create financial and regulatory and legal issues and liabilities for the company. And you certainly don't wanna destroy client or customer relationships over AI data leaks. You want to trust your own use of AI well enough that your clients, your customers, vendors, anyone you interact with as a business or individually in your in your role in your business that others trust you with their data and how you're using AI as much as you trust it yourself. And that's really what it comes down to. And so when we talk about, next steps, no one's asked any questions yet. Now's a great time to hop into the q and a and check your quest you know, any questions. Reminder that we have the resource library. In the resource library, you'll notice we have, an AI landing zone and a cybersecurity landing zone. Those are, really a collation or gathering of all of our library content around those topics. Some of the content will be in both the landing zones if we're talking about AI security in particular. It's a great way to hop in white papers, tools, resources, blog posts, other other web event and session recordings. It's all there. If you have a question and there's information that's not there, feel free to ask. We're happy to look into it, do some research, provide you the information, and then share it in the library, as well. If it's specific to your environment, the best thing to do is to request an introductory call to discuss your AI program, your current status, your needs, and plans going forward with one of our cloud advisers. And for that, you can use the bit.ly CG intro call URL that's on the screen. You can hit our website, or you can use the Bitly code, the QR code on the right hand side of the screen, and that'll bring you to, a screen where you can actually sign up and pick a time that works well for you to start meeting with our cloud advisers. Those conversations are you know, the first call is fairly brief. It's introductory. It's getting a sense of lay of the land. We'll talk about options, how you might wanna proceed. We'll have conversations with you and help you figure out the right path. We'll let you know where we can help and where you might need some different resources than we provide, and sort of help you come up with your road map and help you execute it as best we can if you wanna work with us in that regard. And so on that note, I'm not seeing anything, any comments in the q and a. I'll check the chat real quick or the chat. I wanna thank you all for joining us here today, whether you're watching live or you've been watching on demand. I'm hoping the information has been helpful and, effective and valuable to you. Feel free to reach out and contact us with any questions. Actually, let me jump back to that slide. You can reach out to us at info@cumulusglobal.com. If you wanna get in touch with me, the best way is by email, afalcon@cumulusglobal.com. I'm not great about answering the phone or checking voice mail, so send me an email. I don't have an assistant or even a virtual assistant, so I'll get your email. If I can't handle it quickly, I will have someone in my team get in touch with you, and I'll get involved as needed as appropriate or as you desire. As I become available and so that's sort of my my promise to you and our promise to you as well, and so I want to thank you for joining everyone have a great day have a great rest of the week. Enjoy the rest of September, and, hopefully, we'll hear from you soon, to talk about your current AI security posture and how you may want to evolve it and be ready for continued use of AI going forward. Oh, actually, before we go oh, just a couple of thank yous. I appreciate that. I just wanted to double check the chat. I got some alerts. So thank you for joining us. Have a great day, and we will catch you soon.