Understanding the risks and realities of deploying Agentic AI inside large organisations
In the current hype cycle, “Agentic AI” is pitched as the ultimate AI silver bullet. The truth is subtler. This piece is an attempt to translate the concept into plain language for executives: what agents actually are, how they differ from models and what has made them feasible now.
Why now: Just recently, a major AI cloud provider announced a multibillion-dollar fiber-infrastructure expansion across three continents to support model training. As hyperscale networks multiply, the glass itself — literal optical fiber — has never borne more strategic weight.
When Edward Snowden exposed the scale of global surveillance, most readers focused on who was watching. I focused on how they were watching — through the glass fiber that carries nearly every signal on Earth. Light escapes a fiber and, with it, privacy. We encrypted the payloads. We forgot the pipe.
Years ago, at a cybersecurity conference in San Francisco, I asked the late Kevin Mitnick a question that became my life’s work: What if a fiber could fight back? He smiled and said, “Then you’d be rewriting the rules.” He was right — but I misjudged what it would take.
I committed the classic innovator’s error: I tried to build a product and a market at the same time. “Physical-layer security” wasn’t a category. There were no RFPs for it, no budgets, no Gartner quadrant. The industry that moves at the speed of light often adopts at the pace of cement. My company spent years proving that you can design fiber links to resist physical tapping intrinsically — treating privacy not as a software promise but as a property of the medium itself.
Why tell this story now?
Because the world just made my early mistake look less like an error and more like an omen. The AI build-out is shoving orders of magnitude more sensitive data through the same glass — between data centers, across clouds, into models, out to devices. We obsess over keys, ciphers, and protocols. We rarely ask whether the conduit itself is defensible. We add locks to the briefcase and then leave the suitcase unzipped.
This is a governance problem as much as a technical one. Boards sign off on risk frameworks that enumerate application and endpoint controls to the nth degree and wave vaguely at “transport.” Cloud contracts imply safety in transit while disclaiming the physical path. Public policy debates circle endlessly around cryptography and lawful access, while the most accessible point of compromise — glass you can bend with a thumbnail — sits outside the conversation.
Securing fiber isn’t only about nation-state actors or movie-plot threats. It’s about scale. When traffic volumes grow, so do incentives to steal at the fattest, quietest choke points. In a world where AI systems are trained on streams of proprietary and personal data, silent interception is no longer a spy novel — it’s a business model.
The answer is not to bury civilization in concrete or wrap every building in conduit worthy of a war zone. It’s to upgrade the baseline: treat physical-layer privacy as a first-class requirement of modern infrastructure. That looks like procurement language that requires intrinsic anti-tap characteristics for new build-outs; measurement and monitoring that detect micro-disturbances at the fiber itself; and funding models that recognize that defending the medium is as essential as encrypting the message.
This isn’t an argument against software security. It’s an argument against magical thinking — the idea that we can keep piling cryptography on top of a medium we don’t control. Defense in depth only works if one of the layers is the thing the data actually travels through.
I did a lot of things wrong for all the right reasons. I underestimated capital needs. I underestimated the communications challenge of explaining a new category to buyers trained to look everywhere but the pipe. I overestimated the willingness of markets to fix a problem before it breaks expensive things in public. But being early is not the same as being wrong. The AI era has synchronized the clock.
We have the engineering patterns to make privacy assured, not assumed — including fiber links designed to resist and reveal physical intrusion. What we need now is the will to set the standard and buy to it. If the world is going to entrust its most sensitive ideas to machines, it should start by securing the glass that carries those ideas.
This blog feeds the narrative that ISPs have legal obligations to protect in-transit data from being eavesdropped: https://harmonyblogs.com/internet-service-provider-liability-2/
About the author
Gary M. Weiner is the founder and CEO of Apriori Network Systems, a company focused on physical-layer privacy in optical networks. He writes from Bedminster, N.J.