For most of its history, security ran on a comfortable assumption. Builders knew they were builders. Things shipped on a schedule. If you looked hard enough, you could find what you were protecting. Inventory the assets, scan them, fix what’s broken, repeat. The cadence was human. The attack surface had the courtesy to hold still long enough to be examined.
That assumption no longer holds. Two shifts broke it, and they feed each other:
- AI expanded who counts as a builder, from the engineer with a copilot to the marketer describing an app in plain English.
- A growing share of what those builders ship isn’t a static workload. It’s an agent that acts, decides, and builds further at machine speed.
The first shift is the one everyone talks about, but most people describe it too narrowly. It isn’t just that developers ship faster. It’s that the definition of “developer” has come apart from both ends. Per JetBrains’ 2026 survey, 90% of developers now regularly use at least one AI coding tool at work. The people you already knew were building are producing more code with less human review than any generation before them, and much of what they ship reflects choices an AI made and no one audited.
On the other end, AI app generators like Lovable, Replit, and Vercel’s v0 have turned building into something a person can do without recognizing all the components being used to run the app. A marketer describes an app in plain language and gets back a working one: database, auth layer, API keys, public endpoint. They didn’t write code. But they just deployed production infrastructure, connected it to real data, and put it on the internet. Vercel’s v0 alone crossed 4 million users in early 2026. These citizen developers aren’t reckless. They’re productive, exactly as their tools designed them to be. But every one expands the attack surface without the mental model that used to come with the ability to deploy. Nothing in the flow tells them there’s anything to secure.
The second shift is quieter and more consequential: the things being built have started to build. An AI agent isn’t a static workload you scan and file away. It holds credentials, calls tools, makes choices about what it finds, and chains actions in sequences no one authorized. The model endpoint a data scientist deployed isn’t a server, it’s a decision-maker with access.
And increasingly, the agent is the builder. Vercel reported this year that 30% of apps on its platform are deployed by AI agents rather than humans. That number will not be smaller a year from now. The population of builders now includes actors that don’t file tickets, provision resources without pausing to consider whether they should, and act at a pace no review process can match. If citizen developers broke security’s ticketing assumption, agentic builders break its clock.
You can’t secure what never told you it exists
The danger with citizen-built apps isn’t that they’re insecure. Plenty of hand-coded systems are worse. It’s that they’re invisible to the people accountable for them. A marketer building on an app generator isn’t trying to evade security; security isn’t a step in the flow. The card gets charged, the app goes live, no ticket gets filed, because nothing in the flow ever signals one is needed. Verizon’s 2026 DBIR ranks shadow AI as the third most common non-malicious insider action, up fourfold in a year. Agents make it worse: they simply act, and the record of what they did sits inside a tool the security team doesn’t know is running.
So the answer has to meet these apps in two places at once: the platform where they’re configured, and the cloud where they run. The platform holds decisions like whether the app shipped with authentication, whether secrets sit in plaintext, whether the workspace enforces SSO. The cloud holds the reachable surface, identities, paths an attacker could take once the app is live. Watch only the platform and you’re guessing. Watch only the cloud and you’re finding these apps by accident, weeks late. See both.
Visibility is the entry, not the answer
There’s a tempting response to all this, and it’s the wrong one: see more, scan more, alert more. The industry has spent a decade proving that more visibility, by itself, produces more noise, not more safety. A team buried under ten thousand findings is not safer than a team buried under a thousand. It’s the same team, more tired, missing the same things.
But discovery has become unavoidable, because the list of places to look is no longer knowable. Cloud security once operated against a stable set: five hyperscalers, a fixed roster, quarterly change. That set has become fifty and rising, counting AI generators, model providers, hosting platforms, and agent frameworks between the builder and the running app. You cannot apply context to what you cannot see. Discovery is the price of entry.
The discipline is judged on what happens next: whether that flood collapses into something a human can act on. Not “there is a vulnerability here,” but “an attacker can reach this one, and here is the path.” Our research keeps finding exposed AI and ML credentials in cloud environments, with no one watching the seam where model access meets cloud misconfiguration. A scanner can find the credential. Only context can tell you it matters. For agent-built systems, context isn’t static. It’s a live question that changes every time the thing acts.
What this asks of security leaders
This discipline has a name: AI AppGen Security. But the name matters less than what it acknowledges: the security models built for a smaller, slower, more human builder population have run out of runway.
The role at the center of all this isn’t disappearing. It’s changing shape. The security leader is no longer the person who inspects what was built and grants or denies passage. There isn’t time. Most of what gets built now never passes through a gate at all, and a growing share of it isn’t built by a person you could stop. No headcount is coming to close the gap. You cannot hire your way to an attack surface that expands every time anyone, human or agent, has an idea.
Platforms such as Orca Security reflect this shift toward context-driven cloud security, helping teams connect assets, identities, exposures, and risks so they can prioritize the issues that create meaningful paths to critical resources rather than treating every finding equally.
The leader becomes the person accountable for enabling every builder without becoming the thing that slows them down. The engineer with a copilot. The marketer on Lovable. The agent acting at machine speed. That’s a harder job than the old one, and a more important one. It means trading control-by-gate for control-by-context: knowing what exists the moment it touches your environment, knowing what’s reachable, and having remediation wired close enough to the work that saying “yes, safely” is faster than saying “wait.”
The companies that pull ahead in the next few years are the ones still building when everyone else has stalled out under the weight of their own attack surface. Building is the activity that creates value now. It has stopped waiting for permission, stopped requiring expertise, and, increasingly, stopped requiring a human. Security’s job is no longer to slow that down until it’s safe. It’s to make the building safe without slowing it down, and to do it at a speed and scale that only works if security itself becomes something that builds.
Related Categories