AI Is Already in Your Enterprise. Now What?

Episode transcript

Transcript provided for accessibility and search. Minor transcription errors may remain.

AI Security MultipliersEpisode 1: AI Is Already in Your Enterprise. Now What?Host: SteveAI Is Already in Your EnterpriseSteve: AI is changing how we work, how we innovate, and how we defend our organizations.But it's also changing the risks.Welcome to AI Security Multipliers, where we talk about artificial intelligence, cybersecurity, and governance from the perspective of the people who actually have to make it work.I'm Steve.And for this first episode, I want to start with a fairly simple observation.AI is already in your enterprise.That's not a prediction.It's probably there right now.The question is: do you know where?And perhaps more importantly, do you know what information it's accessing?Who Am I, and Why Another AI Podcast?Before we get into that, I should probably tell you a little bit about me, and why I've decided the world apparently needs another podcast about AI.I'm Steve.I'm originally from Britain, although I've lived in the United States for more than twenty years.Before that, and throughout my career, I've had the opportunity to work in all sorts of places around the world, with different companies, different cultures, and very different approaches to technology.I've been working in the IT industry for more than thirty years, and for roughly the last twenty, I've been leading cybersecurity organizations.I'm a CISSP, a CISM, and most recently, AAISM certified in artificial intelligence security management.So I've spent quite a lot of time around technology and cybersecurity.Long enough to have seen several things that were apparently going to change absolutely everything.Some of them did.Some of them, rather less so.I've worked with large technology transformations, cybersecurity programmes, cloud, infrastructure, data, and now enterprise AI.I've sat in the CIO seat.I've sat in the CISO seat.I've worked with executives and boards.And I've spent an extraordinary number of hours in meetings discussing technology strategy.Thirty years in IT does teach you quite a lot.Although, if I'm being fair, it also teaches you that some of the industry's brilliant new ideas are remarkably similar to the brilliant new ideas we had fifteen years ago, just with a better PowerPoint deck.AI, though, does feel different.Not because I believe every claim being made about it.I don't.And we'll certainly get into some of those on this podcast.What interests me is the speed.The accessibility.And the fact that AI doesn't fit neatly into one of our traditional technology boxes.It's a technology issue.A cybersecurity issue.A data issue.A privacy issue.A governance issue.And increasingly, a business strategy issue.Which means everybody owns a little bit of it, and there's always the danger that nobody actually owns it.That's the interesting bit.And that's really why I started AI Security Multipliers.I don't want to spend thirty minutes every week breathlessly discussing whatever AI product somebody announced yesterday.There are plenty of places to get that.I'm much more interested in what happens on Monday morning.You walk into work.Your employees are using AI.Your developers are using AI.Your security team is using AI.Every software vendor you have seems to have added the word "AI" to their product.Your executives want an AI strategy.And eventually somebody turns to the CIO or the CISO and asks:"So, are we actually okay with all of this?"Usually about six months after everybody started using it.That's the conversation I want to have.AI Didn't Wait for ITThink about how enterprise technology traditionally arrived.Somebody identified a requirement.There was a business case.Maybe an RFP.Procurement got involved.Architecture reviewed it.Security reviewed it.Legal reviewed the contract.Infrastructure got involved.Then there was an implementation project.Testing.Training.Deployment.And eventually, employees were allowed to use the technology.It could take months.Sometimes years.And occasionally, by the time we'd finished implementing it, the original business requirement had changed anyway.But that's another episode.AI didn't arrive like that.An employee can open a browser and have access to incredibly powerful AI capabilities in about thirty seconds.No implementation project.No architecture review.No security assessment.No change request.No training.No eighteen-month transformation programme with a very impressive name.Just open a browser.And that changes things.Because in many organizations, IT didn't introduce AI to the employees.The employees introduced AI to the organization.Think about that for a moment.We normally control technology adoption by controlling access to the technology.With generative AI, that assumption disappeared almost overnight.And if we're still trying to govern AI entirely through processes designed for traditional enterprise software, we're probably going to struggle.Or we'll create a thirty-page AI policy, put it on the intranet, and confidently assume everyone has read it.Which, historically, has always worked brilliantly.The Employee's PerspectiveLet's look at this from the employee's perspective.It's Monday morning.You've got a meeting at ten o'clock.At nine fifteen, somebody sends you a thirty-page document."Can you review this before the meeting?"Of course.Because apparently you weren't doing anything else this morning.You've got forty-five minutes.So you open ChatGPT.Or Claude.Or Gemini.You upload the document and say:"Summarize this for me. Give me the five most important points, and tell me what I should challenge in the meeting."Thirty seconds later, done.And that's incredibly useful.Now let's change one thing.That document contains confidential customer information.Suddenly, we've got a cybersecurity problem.But here's where I think security leaders have to be careful.I don't think that employee woke up that morning thinking:"You know what? Today I'm going to violate the company's cybersecurity policy."That's not what happened.They had a problem.AI solved it.Quickly.So yes, as a CISO, I need to ask:"What was in that document?"Absolutely.But I also need to ask another question.Why did the employee use ChatGPT in the first place?Because if my answer is simply:"Don't do that again."I haven't solved their problem.I've solved my problem.And there's a difference.What I really want is to give that employee a safe way to get exactly the same benefit.Because here's what I've learned.If the secure option takes fifteen minutes, three approvals, and a service ticket, and the unapproved option takes thirty seconds, we shouldn't be terribly surprised by what happens next.That's where I think cybersecurity gets interesting.We're not there simply to say no.We're there to understand what the organization is trying to accomplish, understand the risk, and help find a safe way to say yes.Shadow AIWe've dealt with versions of this problem before.Shadow IT certainly isn't new.Employees have been finding their own applications, cloud services, file sharing tools, messaging platforms, and productivity software for years.Usually because the approved solution didn't quite do what they needed it to do.Or occasionally, because nobody could remember how to request the approved one.AI is really an evolution of that problem.But there are some important differences.First, the barrier to entry is almost zero.Second, the capability is enormous.And third, AI interacts with information in a fundamentally different way.You're not simply storing a document somewhere.You might be asking a model to analyze it.Summarize it.Compare it with other information.Rewrite it.Generate something completely new from it.Or increasingly, take action based on it.That's a rather different proposition.So yes, I think organizations need to understand their shadow AI environment.What tools are people using?What data is going into them?What accounts are being created?Are employees using free consumer services?Are they using enterprise versions?Do we actually know?But here's where I'd be careful.If our entire AI security strategy becomes "find it and block it," we've missed something important.Because if hundreds or thousands of employees are independently finding ways to use generative AI, they're telling us something.There's value there.They're solving problems.They're saving time.They're finding easier ways to do things.Now, some of those ways might make the security team slightly nervous.Fair enough.That's our job.But before we block the technology, we should understand the demand.Because the real question isn't:"How do we stop people using AI?"It's:"How do we let them use it safely?"And those are very different questions.AI Has a Habit of Finding Your Old ProblemsI've been involved with Microsoft 365 Copilot in a large enterprise environment with more than a thousand active users.And one of the most interesting things about that experience was discovering that AI has a rather impressive ability to expose problems you've been happily living with for years.Permissions are a good example.Imagine an employee has access to a SharePoint site.Technically, they're authorized.Maybe that permission was granted three years ago.They've changed jobs since then.They've forgotten about the site.The site owner has forgotten about them.And nobody has looked at the security group since approximately the invention of the iPhone.I'm exaggerating.Slightly.Historically, that might not have created an obvious problem.The employee would need to remember the site existed.Find it.Navigate through folders.Open documents.Search around.And eventually figure out whether there was anything useful there.Now introduce an AI assistant capable of finding and summarizing information across the environment.Suddenly, information that was technically accessible, but practically quite difficult to discover, becomes remarkably easy to find.And this is an important distinction.The AI didn't create the permission problem. It made the existing problem useful. And potentially dangerous.That's why organizations deploying enterprise AI suddenly find themselves talking about things like data classification.Identity governance.SharePoint permissions.Retention.Information protection.Data loss prevention.All those things we've been meaning to tidy up for years.AI has essentially walked into the room, switched the light on, and said:"You might want to have a look at this."And, inconveniently, it's right.Start With the DataSo when somebody asks me:"Where should we start with AI security?"I've become increasingly convinced that one of the best answers is: start with your data.Not the model.Not the clever prompt.Not the latest AI product.The data.What information do you have?Where is it?Who owns it?Who can access it?How is it classified?How long do you retain it?Can employees share it externally?Can applications access it?And now, can AI access it?These aren't particularly glamorous questions.Nobody is going to stand on stage at a technology conference and announce the exciting new world of SharePoint permission cleanup.Although, to be fair, I'd probably attend that session.Because this stuff matters.AI dramatically increases our ability to discover, combine, summarize, and act on information.Which means weak data governance becomes much more visible.If everybody in the organization can access something they shouldn't, AI can potentially make that mistake much easier to exploit.If sensitive data isn't classified, AI doesn't magically understand what you intended.And if information should have been deleted five years ago, but wasn't, well, AI may suddenly make it useful again.Here's another way of thinking about it.If your organization has perfectly classified data, pristine permissions, clear ownership, and nobody has access to anything they shouldn't, congratulations.I'd quite like to visit.Because I've never seen one.The rest of us probably have some work to do.So before we get too excited about the model, look at the data.Now the AI Needs an IdentityThen we get to identity.And I think this is going to become one of the most important cybersecurity conversations around AI.We've spent years improving human identity.Multifactor authentication.Single sign on.Conditional access.Privileged access management.Least privilege.Identity governance.We've made quite a bit of progress.And just as we're getting comfortable with all of that, we've decided to create autonomous AI agents.Excellent timing.Because an agent may need its own identity.It may need credentials.It may need access to an API.A database.Microsoft Graph.SharePoint.ServiceNow.Salesforce.GitHub.Azure.AWS.Or some wonderfully complicated combination of all of them.And this is where it gets interesting.What can the agent access?What can it change?Who approved those permissions?How long do those permissions exist?Can another agent instruct it?Can it delegate authority?Can it initiate a financial transaction?Can it change production data?Can it send an email?Create an account?Disable one?And what happens when it does something we didn't expect?Not if.When.Because software has been doing unexpected things for as long as we've been writing software.Adding artificial intelligence hasn't suddenly made that problem disappear.So, can we stop the agent?Can we revoke its access?Can we see what it's doing?And afterwards, can we reconstruct exactly what it did?Those sound like AI questions.But they're also identity questions.Security architecture questions.Audit questions.Risk questions.And governance questions.The boundaries are starting to disappear.And I think that's going to force some very interesting conversations between teams that historically haven't always worked particularly closely together.Please Don't Create Another CommitteeThere's another mistake I think we need to avoid.Creating an entirely separate governance universe for AI.New committee.New process.New risk framework.New approval board.New terminology.And, obviously, a new form.Because nothing says innovation quite like another approval form.Now, to be clear, AI does introduce genuinely new risks.I'm not suggesting we just throw it into the existing process and hope for the best.But organizations already govern things.Technology.Cybersecurity.Privacy.Data.Third parties.Software development.Architecture.Legal risk.Compliance.Business continuity.So rather than immediately building another island, I'd start by asking where AI fits into what we already have.What's genuinely new?What existing controls still work?Where are the gaps?And where does the risk actually justify additional governance?Because if every use of AI requires a six-week approval process, people will find another way.And if every use gets automatically approved because we've decided we're an innovative company now, well, we're not governing anything.The answer, I think, is proportionality.An employee using an approved AI tool to rewrite public marketing content is one level of risk.Using AI to make decisions about employees is another.An autonomous agent with privileged access to production systems?That's something else entirely.Those shouldn't go through identical governance.That sounds obvious when you say it out loud.But large organizations have an extraordinary ability to make obvious things complicated.So govern the risk.Not the buzzword.Security Needs to Understand WhyThis is where cybersecurity can sometimes get itself into trouble.We see the risk before we see the opportunity.That's understandable.It's part of the job.Somebody has to be the person in the room asking:"Hang on. What happens if this goes wrong?"And quite often, that's us.But if the business sees enormous value in something, and security walks into the room with a PowerPoint containing twenty-seven reasons why they can't do it, we shouldn't be terribly surprised when people stop inviting us to the early meetings.I want security in the room earlier.Before the technology is selected.Before the contract is signed.And preferably, before somebody has promised the board that AI is going to transform the entire company by next Tuesday.To earn that seat, though, security needs to understand the business case.What are we actually trying to accomplish?What's the value?What happens if we don't do it?What alternatives exist?And then, what are the risks?Now we're having a useful conversation.Maybe we can reduce the risk.Maybe we change the architecture.Maybe we restrict the data.Maybe we put a human approval step into the process.Maybe we monitor the activity.Maybe there's a safer model.A safer deployment approach.Or perhaps the residual risk really is unacceptable.That happens too.But now we're making a business decision informed by cybersecurity.Not a cybersecurity decision imposed on the business.That's an important difference.Security shouldn't be the department of no.But neither should we become the department of:"Well, everyone else is doing it, so I suppose it'll be fine."There's quite a lot of space between those two positions.That's where we need to operate.Write a Policy People Might Actually ReadOrganizations also need an acceptable use policy for AI.But please, write one that normal people can actually understand.Employees shouldn't need a law degree, a cybersecurity certification, and a quiet afternoon to figure out whether they're allowed to ask Copilot to summarize a document.The policy should answer practical questions.What AI tools am I allowed to use?What information can I put into them?What information can't I put into them?Can I use AI-generated content externally?Do I need to verify the output?What about source code?Customer data?Personal information?Confidential information?And what do I do if I'm not sure?That last one matters.Because uncertainty is inevitable.The technology is moving far too quickly for a policy to anticipate every possible use case.So give people somewhere safe to ask.And make the answer reasonably quick.If somebody asks a sensible question on Monday and receives an answer three weeks later, they've probably already found another solution.A good policy gives people principles.Not just prohibitions.And perhaps most importantly, don't measure success by whether you've published the policy.Measure it by whether anybody understands it.Those are also not always the same thing.AI Is Also Working for the DefendersOf course, there's another side to all of this.AI isn't just something cybersecurity has to secure.It's becoming part of cybersecurity itself.And this is one of the areas I'm genuinely excited about.Security teams deal with enormous amounts of information.Alerts.Logs.Endpoints.Network activity.Cloud environments.Identity events.Threat intelligence.Vulnerabilities.Application telemetry.The list goes on.An analyst might have to move between five or six different systems just to understand what happened during one incident.And anyone who's worked in a security operations center knows the problem is rarely a shortage of alerts.We've become extremely good at producing those.Understanding which ones actually matter?That's a bit harder.AI can help.It can summarize incidents.Explain scripts.Construct queries.Correlate information.Suggest investigation paths.Translate technical information.And help analysts understand technologies they don't work with every day.That's real value.But then we come back to our old friend, governance.How much authority are we willing to give the AI?Would I let AI summarize a security incident?Absolutely.Would I let it recommend the next investigation step?Sure.Would I let it isolate a laptop?Maybe.Would I let it disable a privileged account?Now I want some context.Would I let it shut down a production environment?Right.Now we're having a very different conversation.There's a spectrum between AI as an assistant and AI as an autonomous decision maker.And I don't think anybody can give you one universal answer for where to draw that line.It depends on the organization.The use case.The consequence of getting it wrong.And how confident we are in the controls around it.The important thing is that we draw the line deliberately.Not discover where it was after something goes wrong.Don't Wait for PerfectOne thing I wouldn't do is wait for the perfect AI governance framework.It doesn't exist.And I'm fairly confident that by the time somebody finishes writing the perfect one, the technology will have changed.Probably twice.So start with some basic principles.Know what tools are being used.Understand the data.Establish accountability.Protect sensitive information.Apply least privilege.Log important activity.Require human oversight where the consequences justify it.Assess higher-risk use cases more carefully.And keep learning.I may change my mind about some aspects of AI governance over the next few years.Actually, I'd be slightly worried if I didn't.Because we're all learning.The technology is changing.The use cases are changing.The threats are changing.So the governance model has to be able to change as well.There's a difference between moving quickly and moving recklessly.There's also a difference between being cautious and being paralyzed.I've seen organizations manage to achieve both at the same time, which is quite an accomplishment.Good governance should help us find the space between those extremes.It should let the organization move with its eyes open.So What Do You Do on Monday?So let's bring this back to Monday morning.You've listened to the podcast.You've had your coffee.You're feeling appropriately concerned about AI.What do you actually do?I'd start with discovery.What AI are we really using?Not what we think we're using.Not just what procurement says we've purchased.What's actually being used?Then look at the use cases.What are people doing with it?What problems are they solving?Then, the data.What information is going into these systems?Then identity.What can these systems and agents access?What can they do?Then governance.Who owns the decisions?Who owns the risk?Who gets to say yes?And who gets to say no?And finally, enablement.Are we giving employees a secure way to get the benefits they're clearly looking for?Because if the approved path is wonderfully secure but completely useless, people will find another path.They always have.AI hasn't changed human nature.The secure path also has to be the useful path.I think that's one of the most important principles we can carry into enterprise AI.Make the safe thing the easy thing.Or at least, don't make it unnecessarily difficult.That's usually a good start.The MultiplierAnd now, this week's Multiplier.One practical thing you can take back to your organization.Find three regular employees.Not technology executives.Not security people.Just three people trying to get their jobs done.And ask them three questions.What AI tools are you using?What do you use them for?And what information do you put into them?Then listen.Don't immediately correct them.Don't turn it into an audit.And whatever you do, don't produce a form.Just listen to the problem they're trying to solve.Because those three conversations might tell you more about your real AI environment than a twenty-page strategy document.And you'll probably discover a use case you hadn't thought about.You may discover a risk you hadn't thought about either.Both are useful.That's this week's Multiplier.ClosingThere's a lot we're going to explore on this podcast.AI agents.Identity.Data.Cybersecurity.Governance.Risk.Architecture.And the technologies that are changing how all of those things fit together.Some episodes will be technical.Some will be more strategic.And sometimes I'll bring in people who know far more about a particular subject than I do.Which shouldn't be terribly difficult.The objective is pretty simple.Useful conversations about artificial intelligence and cybersecurity without getting lost in either the hype or the fear.I don't know exactly where AI takes us over the next five or ten years.I'm not sure anybody does.And anyone who tells you they know with absolute certainty, well, I'd probably ask a few more questions.What I do believe is that the organizations that succeed with AI won't be the ones that eliminate every risk.That's impossible.And they won't be the ones that simply ignore the risks and move as quickly as possible.They'll be the organizations that learn how to innovate while understanding the risks they're taking.Good technology.Good security.Good governance.And good people making sensible decisions.That's really what AI Security Multipliers is about.I'm Steve.Thanks for spending some time with me.And I'll see you next time.