Research and advisory firm Gartner predicts that agentic AI will be in 33% of enterprise software applications and enable autonomous decision making for 15% of day-to-day work by 2028. As enterprises work toward that future, leaders must consider whether existing cloud infrastructure is ready for that influx of AI agents.  “Ultimately, they are run, hosted, and are accessed across hybrid cloud environments,” says Nataraj Nagaratnam, IBM fellow and CTO of cloud security at technology and consulting company IBM. “You can protect your agentic [AI], but if you leave your front door open at the infrastructure level, whether it is on-prem, private cloud, or public cloud … the threat and risk increases.” InformationWeek spoke with Nagaratnam and two other experts in cloud security and AI to understand why a secure cloud infrastructure matters and what enterprises can be doing to ensure they have that foundation in place as agentic AI use cases ramp up.  Security and Risk Considerations  The security and risk concerns of adopting agentic AI are not entirely unfamiliar to organizations. When organizations first looked at moving to the cloud, security, legacy tech debt, and potential data leakage were big pieces of the puzzle.  “All the same principles end up being true, just when you move to an agentic-based environment, every possible exposure or weakness in that infrastructure becomes more vivid,” Matt Hobbs, cloud, engineering, data, and AI leader at professional services network PwC, tells InformationWeek.  Related:For as novel and exciting as agentic AI feels, security and risk management of this technology starts with the basics. “Have you done the basic hygiene?” Nagaratnam asks. “Do you have enough authentication in place?” Data is everything in the world of AI. It fuels AI agents, and it is a precious enterprise resource that carries a lot of risk. That risk isn’t new, but it does grow with agentic AI.  “It's not only the structured data that traditionally we have dealt with but [also] the explosion of unstructured data and content that GenAI and therefore the agentic era is able to tap into,” Nagaratnam points out.  AI agents add not only the risk of exposing that data, but also the potential for malicious action. “Can I get this agent to reveal information it's not supposed to reveal? Can I compromise it? Can I take advantage or inject malicious code?” Nagaratnam asks. Enterprise leaders also need to think about the compliance dimensions of introducing agentic AI. “The agents and the system need to be compliant, but you inherit the compliance of that underlying … cloud infrastructure,” Nagaratnam says.  Related:The Right Stakeholders Any organization that has embarked on its AI journey likely already realizes the necessity of involving multiple stakeholders from across the business. CIOs, CTOs, and CISOs -- people already immersed in cloud security -- are natural leaders for the adoption of agentic AI. Legal and regulatory experts also have a place in these internal conversations around cloud infrastructure and embracing AI.  With the advent of agentic AI, it can also be helpful to involve the people who would be working with AI agents. “I would actually grab the people that are in the weeds right now doing the job that you're trying to create some automation around,” says Alexander Hogancamp, director of AI and automation at RTS Labs, an enterprise AI consulting company.  Involving these people can help enterprises identify use cases, recognize potential risks, and better understand how agentic AI can improve and automate workflows.  The AI space moves at a rapid clip -- as fast as a tidal wave, racehorse, rocket ship, choose your simile -- and just keeping up with the onslaught of developments is its own challenge. Setting up an AI working group can empower organizations to stay abreast of everything happening in AI. They can dedicate working hours to exploring advancements in AI and regularly meet to talk about what this means for their teams, their infrastructure, and their business overall.  Related:“These are hobbyists, people with passion,” says Hogancamp. “Identifying those resources early is really, really valuable.” Building an internal team is critical, but no enterprise is an island in the world of agentic AI. Almost certainly, companies will be working with external vendors that need to be a part of the conversation.  Cloud providers, AI model providers, and AI platform providers are all involved in an enterprise’s agentic AI journey. Each of these players needs to undergo third-party risk assessment. What data do they have access to? How are their models trained? What security protocols and frameworks are in place? What potential compliance risks do they introduce?  Getting Ready for Agentic AI  The speed at which AI is moving is challenging for businesses. How can they keep up while still managing the security risks? Striking that balance is hard, but Hobbs encourages businesses to find a path forward rather than waiting indefinitely. “If you froze all innovation right now and said, ‘What we have is what we're going to have for the next 10 years,’ you'd still spend the next 10 years ingesting, adopting, retrofitting your business, he says.  Rather than waiting indefinitely, organizations can accept that there will be a learning curve for agentic AI.  Each company will have to determine its own level of readiness for agentic AI. And cloud native organizations may have a leg up.  “If you think of cloud native organizations that started with a modern infrastructure for how they host things, they then built a modern data environment on top of it. They built role-based security in and around API access,” Hobbs explains. “You're in a lot more prepared spot because you know how to extend that modern infrastructure into an agentic infrastructure. Organizations that are largely operating with an on-prem infrastructure and haven’t tackled modernizing cloud infrastructure likely have more work ahead of adopting agentic AI.  As enterprise teams assess their infrastructure ahead of agentic AI deployment, technical debt will be an important consideration. “If you haven’t addressed the technical debt that exists within the environment you're going to be moving very, very slow in comparison,” Hobbs warns.  So, you feel that you are ready to start capturing the value of agentic AI. Where do you begin?  “Don't start with a multi-agent network on your first use case,” Hogancamp recommends. “If you try to jump right into agents do everything now and not do anything different, then you're probably going to have a bad time.” Enterprises need to develop the ability to observe and audit AI agents. “The more you allow the agent to do, the more substantially complex the decision tree can really be,” says Hogancamp.  As AI agents become more capable, enterprise leaders need to think of them like they would an employee.  “You'd have to look at it as just the same as if you had an employee in your organization without the appropriate guidance, parameters, policy approaches, good judgment considerations,” says Hobbs. “If you have things that are exposed internally and you start to build agents that go and interrogate within your environment and leverage data that they should not be, you could be violating regulation. You're certainly violating your own policies. You could be violating the agreement that you have with your customers.” Once enterprises find success with monitoring, testing, and validating a single agent, they can begin to add more.  Robust logging, tracing, and monitoring are essential as AI agents act autonomously, making decisions that impact business outcomes. And as more and more agents are integrated into enterprise workflows -- ingesting sensitive data as they work -- enterprise leaders will need increasingly automated security to continuously monitor them in their cloud infrastructure.  “Gone are the days where a CISO gives us a set of policies and controls and says [you] should do it. Because it becomes hard for developers to even understand and interpret. So, security automation is at the core of solving this,” says Nagaratnam.  As agentic AI use cases take off, executives and boards are going to want to see its value, and Hobbs is seeing a spike in conversations around measuring that ROI.  “Is it efficiency in a process and reducing cost and pushing it to more AI? That's a different set of measurements. Is it general productivity? That's a different set of measurement,” he says.  Without a secure cloud foundation, enterprises will likely struggle to capture the ROI they are chasing. “We need to modernize data platforms. We need to modernize our security landscape. We need understand how we're doing master data management better so that [we] can take advantage and drive faster speed in the adoption of an agentic workforce or any AI trajectory,” says Hobbs.