What Established Security Programs Must Have Before Implementing AI

Enterprise security leaders face mounting pressure to adopt AI-powered security products, yet many lack the processes to evaluate what they are purchasing or control what those tools can reach. According to the World Economic Forum, roughly one-third of organizations have no process for assessing AI tool security before deployment, even as 87% of leaders identify AI-related vulnerabilities as the fastest-growing cyber risk. The National Institute of Standards and Technology (NIST) warns that third-party generative AI tools can introduce privacy and information-security risks based on the data and systems they access.

Survey Data Points to a Governance Gap

Research from the Cloud Security Alliance (CSA) found that 53% of organizations have had AI agents exceed their intended permissions, and 47% experienced a security incident involving an AI agent in the past year. ISACA reports that 56% of digital trust professionals do not know how quickly they could halt an AI system during a security incident. These figures suggest adoption is outpacing oversight.

A Practitioner Framework for AI Deployment

Mark Alvarado, CISO at Academy Sports + Outdoors, discussed these challenges on the AI in Business podcast with Emerj's Yolandi de Weerdt. Alvarado brought more than 20 years of cybersecurity, risk, and compliance experience, including prior roles at Ovintiv and Norton Rose Fulbright. At Academy Sports + Outdoors, he established an incident response framework that reduced recovery time by 90%.

Alvarado outlined three priorities for security leaders. First, problem-led evaluation: define the business problem, possible solutions, and true cost of ownership before granting any AI tool access to enterprise identities, data, or systems. He also recommends establishing a shared definition of "AI" during purchasing discussions, since buyers and vendors may use the term differently.

Second, identity, data, and network visibility as deployment foundations. Teams should know which account an AI system uses and be able to disable that user ID quickly, understand where critical data sits and how it is classified, and see data in transit to inspect for malicious payloads and terminate connections if needed. "If a security program wasn't already buttoned up," Alvarado said, AI "has only made your program exponentially worse."

Third, response thresholds tied to material business risk. Alvarado distinguishes asking a user to confirm unusual activity from disabling an identity or terminating a network connection. When a pattern is egregious enough that the organization cannot afford to get it wrong, he acts first and investigates second.

What to Watch

Alvarado's central argument is that organizations should start with a defined business problem, then build controls that turn an AI-generated signal into a defensible security decision. The materiality threshold differs by company, and each organization must determine its own. As AI agents proliferate and regulatory scrutiny intensifies, the gap between adoption speed and governance maturity will likely determine which security programs hold up under pressure.