Shadow AI Is Already Inside Your Enterprise | Anthony Russell
16 minAudio
Send us Fan Mail Your engineers may be using AI agents your security team doesn't even know exist. Those agents could be accessing sensitive data, handling company code, or operating with credentials nobody is monitoring. Welcome to the growing problem of Shadow AI. In this episode of ClearTech Loop, Jo Peterson sits down with Anthony Russell, founder and principal security researcher at Squid Security, to examine what happens when enterprise AI adoption moves faster than security and governance. Anthony challenges the way organizations are approaching AI oversight, arguing that some governance committees have become little more than security theater. Meanwhile, engineers are finding ways to use AI tools outside approved environments, creating risks that traditional identity and access management weren't designed to address. The conversation also tackles a question every CISO should be asking: If you walked into an organization with widespread AI adoption but no inventory, no agent-specific controls, and limited monitoring, what would you do first? Anthony's answer is practical. Before you can secure your AI workforce, you need to know what exists, who owns it, and what it can access. In This Episode Shadow AI: Why engineers are adopting AI outside approved enterprise environments. AI Agent Security: How short-lived credentials, scoped permissions, and least privilege reduce exposure. Governance vs. Security Theater: Why committees and policies alone aren't enough. Intellectual Property Risk: How AI-assisted development and unmanaged workflows can expose proprietary company information. The Agentic Workforce: What executives misunderstand about AI capabilities and security risks. Human Oversight: Why AI agents need boundaries, accountability, and appropriate controls. The First 30 Days: Why a complete AI inventory should be every CISO's starting point. Worth Reading + Watching Squid Security: Explore Anthony Russell's cybersecurity research and practical security insights. https://squidhacker.com/ NIST AI Risk Management Framework: Guidance for identifying, managing, and governing enterprise AI risks. https://www.nist.gov/itl/ai-risk-management-framework OWASP Agentic AI Threats and Mitigations: Research into the emerging security threats associated with autonomous AI agents. https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/ Stay in the Loop Subscribe to ClearTech Loop for direct conversations with technology leaders about AI, cybersecurity, cloud, and the issues changing enterprise technology right now. YouTube: https://www.youtube.com/channel/UCR7xPyjXPgQJKMBvnjoGBQw Jo's LinkedIn Newsletter: https://www.linkedin.com/newsletters/7346174860760416256/ See you in the Loop. ▶ Watch on YouTube: https://www.youtube.com/@ClearTechResearch/videos 📰 Subscribe to the Newsletter: https://www.linkedin.com/newsletters/7346174860760416256/
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
SPEAKER_00Hey everyone, thank you so much for joining this week's edition of ClearTech Loop. I'm Joe Peterson. I'm the CIO of Clarify 360 and the chief analyst at ClearTech Research. And I'm here today with Anthony Russell. Hi, Anthony.
SPEAKER_01Hey, how's it going?
SPEAKER_00Good. If you're not following Anthony on LinkedIn, you probably should be. Anthony is the founder and principal security researcher for Squid Security.
SPEAKER_01Yeah. Hi, nice to uh nice to meet you. Thank you so much for having me on and bringing me to speak about this.
SPEAKER_00Oh, we're glad to have you. So thanks. So let's get right down to it. We asked three questions, as everyone knows, on this webcast, and today's no exception. Um, one of the ones that's on everybody's mind right now is the fact that AI agents hold delegated credentials, they make authorization decisions in real time, and they can spawn sub-agents with their own permission sets. Traditional IAM and PAM tools weren't built for this.
SPEAKER_00So, what is a realistic path, the least privilege access for AI agents in a large enterprise today?
SPEAKER_01Yeah, uh, so this is uh excellent question. It's very relevant to the current environment, specifically for enterprises. Um I think there's a few things that enterprises are currently doing that are anti-patterns. Uh, but these aren't anti-patterns specifically related to uh AI consumption. They're anti-patterns uh as a whole, as how their their organization functions. So um something that should not come to a surprise to anybody is that my recommendations for uh for access management for uh AI models should be you should be providing short-lived tokens.
SPEAKER_01The tokens should be scoped, and uh you shouldn't be sharing them amongst each other. You should have them specific for each thing. Uh I mean these are these are things that we have been doing for years in in software. Uh it's not new. We don't treat, we shouldn't be treating these uh AI any different. We should be considering uh least privilege for for most of them, uh, for everything you're doing. So, what does that mean? Um that means uh we want to bind every action to a human or an org so we know when something happens, we we know who to speak to immediately.
SPEAKER_01Uh, we don't want to so if you have an incident or something happens, something goes wrong, one of your LLMs goes off the rails and starts uh starts deleting things, you want to know who to speak to. And if you have a giant org of 30,000 engineers or whatever, uh it's gonna be a challenge to track down that person. So make sure that uh the tokens that you're creating um are well documented uh for who they belong to, the purpose, the data class, the expiry. Make sure that they're getting revoked uh in a uh reasonable time as soon as the task is done.
SPEAKER_01You should be revoking these tokens. Um there's other things that we can go into as well uh that that would help prevent this. But does that uh start to answer your question?
SPEAKER_00Yeah, that's good. Um I I think that it's my sense that most organizations have more agents than they know they have.
SPEAKER_01Oh, yeah, absolutely.
SPEAKER_00Um, and that's a little scary in itself. And it it actually, you know, brings me to the next question. And look, I want to applaud every company that has put together a dedicated AI governance committee because I think that's a step in the right direction. And in fact, 50% of large enterprises already have one. But the question that I keep coming back to in my own head is is that structure actually working or is it creating the appearance of governance without the substance? What do you think?
SPEAKER_01Yeah, um as you said, I'm happy that companies are attempting to govern this. They're they're they're taking a step in that direction, right? That is uh the the first the first way to uh to fix a problem is to identify the problem, right? So uh it governance and committees are good. Let's start with that. However, uh there's some problems. Uh the governance committees that I have personally seen, my my personal experience is that they tend to be security theater.
SPEAKER_01Uh, they tend to be rubber stamp, uh, they tend to be run by people that have rarely, if ever, touched production controls and production environments or uh had any hands-on experience engineering. Um, they tend to make decisions based on fear that the industry has put out uh in the media instead of listening to experts at their very own companies. Um there are ways to govern, and uh there are ways to implement these systems in a safe manner.
SPEAKER_01Uh, I recently spoke at DEF CON about uh creating constrained agents. Uh so um this this is something that's at the forefront of our industry. We are thinking about these things. However, uh large enterprises are doing what they do best. They're putting people, they're making people responsible, uh, but they're not really they're not really solving problems. Uh these committees will work much better uh when there's clear risk tiers and when there's clear criteria on when to kill uh when to kill something.
SPEAKER_01Okay, that that's probably a big part of this. Uh we need mandatory control baselines like data boundaries and humans in the loop, um, all types of things. It need to be measurable KPIs and exception tracking, and uh we need to be looking at how engineers are uh using AI realistically, like the shadow AI, you know, the the ones that are not being tracked by by the corporate office or by by leadership. You know, how are how are engineers running this on their local machines?
SPEAKER_01So there's a lot of room for growth. We'll say that. Uh but the the long and short of it is that um shadow AI, overprivileged agents, and incidents tend to be the the biggest issues. Uh, and I'm not quite sure if um the governance we currently have in place is solving these problems.
SPEAKER_00Well, and let's stay there for a minute because I feel like and what I'm hearing customers say is that shadow AI is the organizational hot potato.
SPEAKER_01Yes, because it's it's awesome. AI is awesome, it's fun to use, it works great. I mean, it it takes it so this problem is is incredibly multifaceted. Okay. So there's uh there's an old thought that um if if there's uh a problem the and the the corporate world is uh trying to solve that problem, they will find a way to solve it, right? And the problem that that they've had for the last decade or so is a lack of talent.
SPEAKER_01We've had a uh not saying that the current engineers aren't talented, I'm saying that there's a lack of engineers, right? Uh and this has caused incredible pay packages and and benefits and and things that companies have done to try and attract that talent. Uh, and we've reached this boiling point now where um people are making insane salaries and companies are giving like free healthcare and all these things uh to try and get the talent, but it's just there's you're you're pulling blood from a stone. There's not enough.
SPEAKER_01There's not enough engineers. So, how do we solve this? Well, AI is a really cheap solution to quite a bit of this. Uh, we can take engineers that were once intern or intermediate level and bump them up quite a bit because we have AI to supplement them now. We can say, hey, take this AI and instead of writing that one function and taking a week to do it, you know, now you can do 10 of them in a day. Um, that's part of the the way they're solving that. The problem is that the AIs aren't ready for that yet.
SPEAKER_01They're not ready for um unconstrained uh decision making. So yeah, it's a multifaceted issue uh that we're that we're facing with this. And shadow AI uh is is a real problem. Uh, and it's a real problem because engineers that are underskilled are using AI to meet the requirements of these companies that want you to have you know 15 years worth of experience when the technology has only been out for four kind of thing, you know? Right. Um yeah.
SPEAKER_00It's it's it's a thing, right? Well, so let's stand still a minute there for a minute. Do you think the non-technical C-suite has a real grasp of the organizational security concerns that come along with the agentic workforce? And if you had a one thing to say to them, what would it be?
SPEAKER_01Do I think let's let's unpack that. Do I think that they have a real understanding of the AI workforce?
SPEAKER_00The organizational security concerns that are coming along with that, having that agentic workforce.
SPEAKER_01Uh no, I don't. I think that most of that is driven by uh the fear, uncertainty, and doubt that the media spreads about AI because that's what gets clicks and views. And um I think that the story is somewhere in between what everybody's perceiving. Uh, there's half of the industry is saying, oh, it can't do that, it'll never replace us. The other half of the industry is saying, oh my gosh, the world is burning down and we need to stop it. And I I think the answer is my my theory is the answer is somewhere in between.
SPEAKER_01Uh, it's definitely not useless. It is useful. I use AI on a regular basis. I use it for uh I was a software engineer for 15 years professionally. So I use it quite heavily, even though I can do it myself. I still use AI because it is there is value there. Uh however, uh it doesn't do everything. I make sure I do heavy code reviews, uh, I do security reviews. It does uh need prompting to make sure that it's gonna put security in place correctly. Um, however, I don't think it's burning the world down either.
SPEAKER_01Uh, it's not to that level yet. It's getting there. I mean, it will be. Uh, we see some pretty uh notable events in the media recently, even uh, with Anthropic admitting that its uh agents were able to break out and get into other companies and very concerning things. But most AI models are not doing that. Uh they're not at that level. Uh, in fact, if they were at that level, I'd be happy because you know, my my pen tests would be even better. Uh you know, I I would find all the things, right?
SPEAKER_01But they're just not there. Yeah, right. So um we are trending in that direction. Uh, I do believe that security is a massive concern for the C-suite. Uh, we should be focusing on these things. Uh, but the approach of check the blocks, check the block governance is is and compliance is the wrong way. That is that is the security model of yesteryear. That is the security model of, you know, humans were the only ones doing the thing, so you know, we'll check the block, it'll probably be okay.
SPEAKER_01And if not, then you know, we'll we'll handle the fallout. Unfortunately, that's not gonna work anymore. Uh, AI is getting really great uh at finding things, it can analyze massive sets of data um incredibly quick, and it's very accurate, it can be very accurate. So uh that's where my research has led me, and what I've been working on uh has been figuring out safe ways to constrain AI models uh to use them in a in a professional manner. So yes, I believe that uh there is room to grow and that C-suite should be concerned about cybersecurity.
SPEAKER_01Um, and I think the way that they begin that is to um make sure they're not doing security theater with their governance.
SPEAKER_00Right. And I, you know, and and so let's paint a picture for a minute. You walk in the door, and you very well could, as the CISO, and you take a look around and you realize ooh, we are way ahead on AI, but we're way behind on security. And what I mean by that is nobody's done an inventory, there's no agent-specific controls, there's no monitoring that you spoke up earlier.
SPEAKER_00What is the single most important thing you're gonna do in the next 30 days?
SPEAKER_01Yes, next 30 days. Uh, so I would focus on building an inventory first. Uh, that would be the most important part. We need to know uh how our company is using AI at every level. Uh, because the reality of it is a lot of companies give their engineering teams, their development teams, uh quite a bit of access, quite a bit of keys to the castle, which is not inherently a bad thing. I mean, I've been an engineer for many years. Uh, the it from what I've seen, the smaller the company, the more access they have, the larger the company, the more segregated it is, you know, somewhere in between.
SPEAKER_01But um, the more access the engineer has, the more capability, the faster they can move. Uh, the more restrictions there are, the slower they move, but the the more safeguards there are for a company. Uh, however, just because somebody has a lower access doesn't mean that they can't hose an environment or share proprietary data unwittingly. Um in the in the last eight months, I've seen not one but two incidents where overseas uh outsourced hired engineers from um from overseas companies, I won't name the companies, but overseas companies, uh, that they unintentionally uploaded uh company intellectual property to GitHub, to public GitHubs.
SPEAKER_01So um, and this was because they were underskilled and they were using AI on personal machines and then committed it back to GitHub and accidentally committed it to their personal GitHub instead of a corporate one. So uh these things are real. Yeah, absolutely real. Uh so next 30 days, uh I would first make an inventory of every AI agent, every every use case of AI. Uh, I would want to know who owns it, uh the purpose behind it, what data it can touch, the credentials and tools it holds, and where it runs. Uh, if you focus on that for the next 30 days, uh, that's gonna unlock quite a bit of capabilities for you as a company.
SPEAKER_01Uh everything else kind of falls in line after that, least privilege, monitoring, risk ratings, all of that. Um, but before you can manage it, you have to measure it. Uh, and that's that's what I would recommend.
SPEAKER_00I think that's great advice for every person that you know hits my inbox on LinkedIn and says, what do I do first? Right? What what and a lot of people are thinking about that because they're in a situation where you know, AI, let's just face it, AI doesn't, it's not binary, it doesn't act the way normal security controls or the things that you would put in place normally um don't respond as well, right?
SPEAKER_01It's a non-deterministic system.
SPEAKER_00Yes, thank you. Much better said. So um, you know, that's great advice. So I really appreciate it. Audience, I hope you appreciate it too. Anthony, thank you for making time to come and visit. Really inspired thinking, and I hope to have you back.
SPEAKER_01Yeah, absolutely. Thank you so much. I appreciate your time.
Send us Fan Mail What if the next insider threat to your organization isn't a person at all? AI agents are moving beyond answering questions. They're accessing systems, executing tasks, requesting credentials, and...
Send us Fan Mail Can AI safely troubleshoot and resolve IT issues without putting your entire endpoint environment at risk? NinjaOne Field CTO Jeff Hunter returns to ClearTech Loop to join host Jo Peterson for a...
Send us Fan Mail What happens when AI becomes capable of making decisions its creators can't fully anticipate or control? AI agents are gaining autonomy, accessing enterprise systems, managing credentials, and making...
Send us Fan Mail Your company has an AI governance committee. Great. But when an AI agent makes a decision that harms your business, who actually has the authority to shut it down? And here's a bigger question: Would...
Send us Fan Mail Your company wouldn't hire an employee without onboarding, supervision, or defined access. So why are AI agents being given responsibilities without those same safeguards? In this episode of ClearTech...