Safeguarding AI in the Cloud: Risks and Best Practices for Secure Innovation

It’s a truth universally acknowledged, at least in IT circles, that when something is marketed as “open,” everyone wants a piece—but no one wants to be left with the security bill. Yet here we are. According to Tenable’s freshly brewed Cloud AI Risk Report 2025, there’s an urgent warning for those who rush to embrace open-source AI and cloud innovation: congratulations, you’re probably making your security engineer twitch.

The Acceleration of AI—and its Risks​

The enterprise world has collectively entered a mad dash to inject AI into every byline, legacy system, and, occasionally, helpdesk chatbot (which may or may not still answer with “Have you tried turning it off and on again?”). Companies are desperate to outpace the competition by adopting AI, and who wouldn’t jump at the chance to use the open-source tools and managed cloud services that make the magic happen?
But as the old saying goes, “With great flexibility comes even greater attack surfaces.” Or maybe I just made that up; regardless, Tenable’s report makes clear that while open-source AI frameworks are the darling of DevOps and data scientists, their openness is a double-edged sword. For all the creativity and speed they unlock, they introduce a delightful buffet of vulnerabilities that IT professionals and CISOs would rather skip.
Just imagine: you’re deploying a shiny new AI model using every open-source package you can get your hands on. It’s agile. It’s cost-effective. The boardroom loves the prototype. But somewhere in those many dependency chains—deep in the roots of that beloved open-source tool—is a vulnerability just waiting for an enterprising threat actor to discover. Ah, the romance of modern business innovation.

Highlights from Tenable’s Research: A Smorgasbord of Exposure​

Tenable put AI workloads under the microscope across the cloud juggernauts—AWS, Azure, and GCP—for a yearlong period through late 2024. The findings? Worrying, even if you like to live dangerously. It turns out that in their race to stand out, enterprises are skipping some critical steps in securing their shiny new toolsets.
Need some stats to keep you up at night? Here you go:

• 28% of AI workloads integrate Scikit-learn, that backpack staple of data scientists everywhere.
• 23% are running with Ollama, another open-source favorite.

Here’s the kicker—these frameworks, especially when run on Unix-based systems (which most are), have sprawling dependency trees. Each new library becomes a potential backdoor, especially if it’s not diligently patched. As Nigel Ng, Tenable’s SVP for APAC & Japan, reminds us, “The openness and flexibility that make these tools powerful also create pathways for attackers.” Almost poetic, if you have a dark sense of humor.
Combine open-source sprawl with the convenience of managed cloud AI services, and things get even spicier. More than half the organizations running on Azure have enabled Cognitive Services, and a quarter of AWS users are playing with SageMaker. Plug-and-play is great until you realize the defaults were set with convenience, not compliance, in mind.

The Elephant in the Data Center: Misconfiguration and Data Exposure​

Here’s where the IT rubber meets the road (and sometimes spins wildly out of control): misconfiguration is rampant. Even in 2024, with decades of cloud “best practices” filling blogs, forums, and the more hopeful corners of Stack Overflow, enterprises routinely leave their AI applications in “wide open” mode.
Tenable’s findings lay out a litany of sins:

• Unpatched vulnerabilities residing happily in cloud-hosted environments;
• Cloud services left in default, “please hack me” configurations;
• Data floating around in the ether, just waiting for curious bots or overly zealous interns to discover it.

If you’re thinking, “Surely, with all the stories of breaches—someone has this under control?”—well, add that to the list of optimistic IT fairy tales, right between “Legacy app modernization is straightforward” and “This year, we won’t need to reset any passwords.”

Why Open Source AI Tools are a Unique Headache​

Let’s drill into the heart of the matter: open-source AI isn’t just software; it’s an ecosystem, a squirming mass of dependencies maintained by thousands of contributors of varying skill, motivation, and free time. When execs hear “open source,” they often picture a bustling bazaar of innovation; they rarely imagine the lonely developer receving a bug report at midnight, or the repo abandoned entirely because its creator has discovered the joys of outdoor living.
This is the secret ingredient of open-source risk: abandoned projects and unvetted code—peppered through production AI systems at the world’s most ambitious companies.
And as AI stacks grow more complex, the odds of a forgotten or poorly maintained dependency increase. It’s no wonder the report zeroes in on the dangers of being “too open for one’s own good.” Not every open door leads to innovation. Sometimes it leads to ransomware.

Managed Cloud Services: Comforting (Until They’re Not)​

You’d think that managed cloud services, with their glossy dashboards and “security by design” promises, would solve all these woes. Not so fast.
Tenable highlights that managed platforms like Azure Cognitive Services or Amazon SageMaker are omnipresent—60% and 25% adoption rates, respectively, among survey respondents. These tools are sold as being “secure by default,” but as anyone who’s ever untangled cloud IAM (Identity and Access Management) settings can tell you, defaults get you only so far.
Real-world consequence? Companies launch pilots at breakneck speed, rarely pausing to harden their cloud configurations. Maybe it’s the pressure from leadership (“Why aren’t we doing what our competitors are doing?”), or maybe it’s a fatal case of “move fast and fix later.” Either way, attackers are delighted to see so many workloads running with far more permissions and far less oversight than they should.

The Illusion of Control: Visibility Is Just as Important as Security​

This is where things get existential. The report makes a not-so-subtle point: without visibility into your cloud AI environments, you’re not just vulnerable; you’re flying blind.
“When you don’t know what’s deployed or how it’s configured,” says Ng, “you risk losing control of your AI environments—and, by extension, the outputs they produce.” For the non-paranoid among us, that’s a chilling thought. Imagine pouring millions into AI research, only to have confidential datasets accidentally exposed, or—worse—having your models subtly manipulated by someone who slipped through a neglected dependency.
Visibility isn’t just a nice-to-have. It’s the foundation for every other control mechanism that security teams labor to implement. Sadly, in the AI gold rush, most are still trying to find their shovels.

The Comedy of (Cloud) Errors: Real-World Implications for IT Professionals​

For the IT pros on the frontline, Tenable’s report reads less like a science fiction yarn and more like a playbook of mistakes they’re asked to fix weekly.
Sure, cloud-native and open-source AI are the future. They promise innovation, lower costs, and—let’s be honest—hours saved scouring proprietary enterprise software documentation. But this flexibility means that defending the perimeter is no longer enough. The perimeter is everywhere: in public Git repos, in the quickstart tutorials pasted directly into production, in configuration files left unencrypted.
The real challenge isn’t just technical—it’s organizational. Patch management for open-source dependencies requires processes as much as tools. Auditing cloud permissions means cross-functional conversations no amount of automation can fully replace. And don’t get started on the headache of aligning AI risk management with endless regulatory frameworks.
There’s comedy here, if you look closely: the more “democratized” AI becomes, the more likely you are to find your future security incident started with a junior dev, a stack overflow snippet, and a perfectly innocent-looking YAML file.

Strengths, Silver Linings, and the Path Forward​

Before you dump your Sklearn and Ollama projects in favor of a monolithic mainframe comeback, know this: Tenable isn’t advocating for an AI-free cloud. Instead, the message is annoyingly pragmatic.
Open-source and managed cloud AI are the core building blocks of our (glorious?) machine learning future. The problem isn’t with the tools, but how we use them. The solution isn’t to lock them away, but to invest in smarter controls, continuous visibility, and—dare I say—an ounce more patience at the design stage.
Here are the silver linings for those who still want to sleep at night:

• By understanding your AI pipeline’s supply chain, you can weed out obvious risks before they bloom.
• Cloud services can be hardened, if you invest the time. There’s nothing quite as satisfying (or risky) as editing IAM policies at 2 AM.
• Cultural change is possible: DevSecOps is more than a buzzword, it’s a survival mechanism. Hint: start by buying your security team more coffee.

And, perhaps most importantly, the biggest strength in the AI arms race will always be human oversight. Automation can catch anomalies, but it can’t replace the curious, skeptical eye of a seasoned IT pro who’s already lived through one or two near-misses.

Hidden Risks: The Unseen (and Uncomfortable) Takeaways​

If there’s a warning hidden in the subtext of Tenable’s report, it’s this: don’t make faith in open source or the cloud your only security plan.

• Dependency chains are the new shadow IT;
• Misconfigurations are as dangerous as any zero-day;
• And “secure by default” covers only a small slice of real-world scenarios.

Relying too heavily on tools—be they open-source frameworks or managed clouds—without hard questions about visibility, configuration, and patch management is like buying the world’s fastest car and never checking the brakes.

What Next? Advice for the Security-Minded (and Those Who Tolerate Them)​

Tenable’s call to action is refreshingly clear: Embrace AI, but do so with your eyes open and your patch management automated.
For those building out AI capabilities:

• Audit your dependency trees before they become dependency forests;
• Don’t trust “default” settings—trust, then verify;
• Treat visibility as a non-negotiable requirement in every design;
• Bake security into your DevOps pipeline so thoroughly it becomes as natural as source control.

And for those in management, let this be your clarion call: let your innovation strategy be guided by both creative ambition and healthy skepticism. AI is indeed shaping the future, but only if it’s built on a foundation you’re constantly testing, hardening, and—above all—understanding.

Closing Thoughts: Surviving the AI Gold Rush without Getting Burned​

The real lesson from Tenable’s deep dive isn’t about doom or dystopia (though both make for catchy headlines). It’s about the perennial IT struggle: balancing innovation with caution, speed with scrutiny.
Open-source AI and cloud at scale are here to stay. The risks, as Tenable has so deftly shown, are not theoretical—they’re already lurking in systems large and small. But they’re also, with the right discipline and a dash of humility, entirely manageable.
After all, anyone can build fast. The winners will be those who build securely, and—if they’re very lucky—without a single meme-worthy security incident to explain at the next all-hands. Stay curious, stay patched, and, above all, stay skeptical—because in the world of AI, the only real constant is change (and maybe a little misconfiguration).

---

Source: Back End News Tenable warns that open-source AI puts cloud security at risk | Back End News