Companies just spent years adapting to cloud, gradually developing new operational models, security controls, and governance frameworks as infrastructure shifted beyond traditional datacenters.
But along came AI, which compressed that timeline dramatically. Enterprises are deploying large language models into customer support, software development, and internal productivity workflows while security teams are still trying to understand the implications. That imbalance is changing both how organizations defend themselves and how attackers identify opportunities.
On a recent episode of the Shared Everything podcast, cybersecurity analyst and longtime practitioner Adrian Sanabria argued that the challenge is that yes, AI introduces new threats, but it also accelerates nearly every existing process on both sides of the security equation.
After more than two decades working as an enterprise defender, penetration tester, analyst, and product leader, Sanabria has watched several major technology shifts reshape enterprise security. None, he says, have moved at the pace of AI. Organizations that once approached new infrastructure cautiously are now integrating generative AI into production environments almost immediately , often before security teams have established meaningful guardrails.
I’ve never seen anything catch on and be absorbed by the enterprise like AI. I kept telling myself, surely nobody’s going to just throw this into an application that faces the public internet. Then all of a sudden we saw these LLM-driven support chatbots, and even the most conservative companies I advise just jumped into this with both feet.
Much of the public discussion around AI security has focused on its ability to discover software vulnerabilities, and for good reason. Models are becoming increasingly capable of identifying exploitable flaws that previously required significant human expertise. But Sanabria argues that vulnerability discovery is only the beginning of the problem.
Enterprises have never struggled to identify security issues but they have struggled to fix them quickly enough, he says.
Security organizations rarely own the infrastructure they protect, meaning they depend on application teams, platform engineers, and operations groups to deploy updates. As AI increases the volume and speed of vulnerability discovery, prioritization becomes even more important than patching everything. Security teams must distinguish between theoretical issues and vulnerabilities that are genuinely likely to be exploited in the wild.
Security doesn’t apply any of these updates. You have to beg, bribe, plead the actual asset owners to update these systems. Most vulnerabilities are meaningless from a security or risk perspective. They’re never going to be used by an attacker. The challenge is figuring out which ones really matter.
The security concerns surrounding large language models extend well beyond traditional software vulnerabilities. Prompt injection, retrieval poisoning, and manipulated training data expose weaknesses that are deeply connected to how today’s models function rather than simple implementation defects. Sanabria points to emerging research suggesting that surprisingly small amounts of carefully crafted content can influence model behavior, creating new opportunities for attackers and marketers alike.
While vendors are developing filters and protective layers to screen prompts before they reach models, many of those defenses rely on AI themselves. The result is an increasingly recursive security architecture in which language models are asked to defend other language models against attacks specifically designed to exploit how they interpret information.
Embedding security into infrastructure design will not eliminate risk, but it does reduce the burden on centralized security teams that are already struggling to keep pace with AI’s rapid adoption.
Some research came out and said as few as 250 documents are enough to influence a model. You could publish 250 blog posts, and that’s enough to ensure a model only says good things about your company or only says bad things about your competitors.
These AI-specific concerns arrive just as enterprises are still addressing older problems that remain far from solved. Supply chain security continues to challenge organizations years after incidents like Equifax and Log4Shell demonstrated how difficult it can be to identify vulnerable components buried inside sprawling software estates. Today, enterprises find themselves balancing two opposing objectives. They must deploy security patches as quickly as possible while becoming more cautious about adopting newly published open source components that could themselves be compromised. That asymmetry continues to favor attackers, who often need only a single overlooked weakness while defenders remain responsible for securing thousands of applications, services, endpoints, cloud environments, and third-party dependencies simultaneously.
Rather than treating security as the exclusive responsibility of security organizations, Sanabria believes ownership needs to move much closer to where infrastructure and software are built. Platform teams, developers, infrastructure operators, and data architects all make design decisions that directly influence security outcomes long before a security review takes place.
