By Luigi Caramico, Founder and CTO, DataKrypto
As enterprise intelligence converges, the infrastructure creating its greatest advantage is also creating a new center of risk.
I grew up in Rome, so I have heard the expression “all roads lead to Rome” my entire life. Most hear it today as a figure of speech, but for the Romans it evokes something much more consequential.
The Roman Empire built one of the great infrastructure networks of the ancient world. Roads connected distant provinces to the center of the empire, allowing armies, officials, messages, goods and people to move across enormous distances with remarkable speed. They helped Rome project power, administer territory, expand trade and maintain communication across an empire that stretched from Britain to North Africa and the Middle East.
Much of the advantage came from reducing the friction of distance. A legion could reach a troubled frontier faster. A message could travel farther. Goods could move more reliably between cities and provinces. Rome became more powerful because its vast territory was more connected.
That connectivity came with consequences. Infrastructure built to move Roman armies could also be used by forces moving against Rome. Trade and travel networks could help disease spread between population centers. Maintaining and defending thousands of miles of roads required resources, and the reach created by the network expanded the territory Rome had to manage and protect.
The Romans had built infrastructure so useful that it inevitably became useful to others as well.
There is an interesting parallel in what is happening with AI today. Enterprises are building a new kind of connective infrastructure, one designed to eliminate informational distance rather than geographic distance. As more data, applications, models and processes connect to AI, we are creating extraordinary new capabilities while changing the shape of what has to be protected.
The Roads Are Converging
For most of the history of enterprise computing, valuable information was fragmented. Customer information lived in CRM systems, financial information in ERP platforms, intellectual property in repositories and databases, and operational knowledge across applications, business units and people.
That fragmentation has always been inefficient, which helps explain the extraordinary appeal of AI. An enterprise model can reason across customer records, financial information, source code, internal documents, telemetry, communications and proprietary data. RAG systems can make large stores of organizational knowledge available during inference. Agents can retrieve information from multiple systems, maintain context, call applications and take action based on what they learn.
In practical terms, AI is reducing the informational distance between parts of the enterprise that previously existed in separate systems and organizational silos. Information that once required several applications, queries and people to assemble can increasingly be brought together in a single computational environment.
That changes security in ways we are only beginning to appreciate.
Fragmentation creates friction for legitimate users, but it also creates friction for an attacker. Valuable information distributed across multiple systems may require different credentials, attack paths and environments to reach. As enterprises connect those systems to AI, some of that separation disappears.
This does not make connectivity a mistake. Connectivity is the reason AI is useful. It does mean that the architecture creating the advantage deserves to be examined from the perspective of what happens when its defenses are breached.
The Advantage and the Exposure Come From the Same Architecture
A pharmaceutical company may connect molecular data, clinical research, experimental results, proprietary models and decades of scientific expertise to AI because insights can emerge from their combination that would be difficult to discover from any one source.
A semiconductor manufacturer may connect process models, manufacturing recipes, equipment telemetry, supplier information and engineering knowledge for the same reason. A financial institution may bring together transaction data, risk models, customer information, market intelligence and institutional expertise.
In each case, the objective is to give AI enough context to understand relationships that were previously distributed across the organization. As that context grows, AI becomes a richer representation of what the enterprise knows and, increasingly, a mechanism through which that knowledge can be put to work.
Security architecture developed in a world where much of that value remained dispersed. AI is steadily concentrating it.
A New Zero Milestone
Near the Roman Forum once stood the Milliarium Aureum, the Golden Milestone, erected by Augustus and traditionally associated with the point from which Rome’s great road network was reckoned. It is sometimes described today as Rome’s zero milestone, although the exact role it played in Roman distance measurement is still debated.
What matters for the analogy is that the stone represented the center of a network. Its value did not reside in the marker itself, but in the system of roads, cities, commerce, communication and power organized around Rome.
AI is beginning to occupy a comparable position inside the enterprise. Data, applications, models, employees and business processes increasingly connect to a common computational environment where information can be combined, interpreted and acted upon.
There is an important difference between the two milestones. The Golden Milestone marked a center, but it did not contain everything that traveled along the roads leading toward Rome. AI increasingly does more than mark the center of the enterprise information network. It can absorb its context, synthesize its knowledge and retain representations of what has been brought together.
As a result, the destination itself becomes valuable in a way the ancient milestone never was.
What Happens When Knowledge Concentrates
The breach of McKinsey’s internal Lilli AI platform earlier this year provided an unusually clear illustration of the implications. Researchers gained access to a system containing millions of chat messages, tens of millions of RAG document chunks, hundreds of thousands of files, system prompts, AI assistants and other elements of the firm’s internal AI environment.
The larger lesson was not simply that a vulnerability existed. The amount and variety of information accessible through the system reflected the purpose for which enterprise AI is being built.
Organizations want AI to understand more of their business. They want models to have greater context and agents to work across more systems. They want AI to draw relationships among information that was previously scattered throughout the enterprise. Every successful step in that direction makes the AI environment more capable, while also increasing the value of what can be reached through it.
This is why I believe the target is changing.
The individual databases, applications and repositories still matter, and attackers will continue to pursue them. But an AI environment that connects many of those resources offers something different: access to the layer where information is being assembled into organizational intelligence.
Protecting the Center
The Romans protected their network with troops, checkpoints, fortifications and control over territory. Modern enterprises have their own defenses in identity, access controls, segmentation, governance, monitoring, attestation and infrastructure isolation. These controls remain essential as AI becomes more deeply connected to enterprise systems.
The question is whether they are sufficient for an environment in which so much value is being concentrated during computation.
Most AI systems still require sensitive information to become plaintext somewhere while it is being used. Model weights are loaded into memory, prompts are processed, and embeddings, activations and intermediate values move through CPUs, memory and accelerators. Enterprises can surround that computation with increasingly sophisticated controls while the information itself remains exposed inside the execution environment.
At DataKrypto, this is the problem we have been working to change. FHEnom for AI™ is designed around the principle that sensitive data and AI models can remain encrypted through computation. Model weights, prompts, embeddings, activations and other sensitive elements of an AI workload remain cryptographically protected through execution, while attestation verifies the trusted environment responsible for key custody.
The practical objective is to reduce what an attacker can gain even if another layer of defense is compromised. A memory snapshot should reveal encrypted information rather than usable intelligence. Access to an accelerator should not provide access to the model weights or data being processed there. An encrypted model taken from an environment should remain unusable outside it.
These properties become more important as enterprises connect more of their information and operations to AI, because there is little reason to believe this trend will reverse. The economic value of bringing information together is too significant.
All Roads Lead to AI
Rome’s roads were among its greatest achievements because they allowed an enormous empire to function as a connected system. They made movement faster, communication more reliable, commerce more efficient and centralized administration possible across vast distances. The same infrastructure inevitably changed the empire’s exposure because anything capable of using those roads could benefit from the connectivity Rome had created.
Enterprise AI presents a modern version of that challenge. We are connecting data, applications, models, people and increasingly autonomous agents because valuable intelligence emerges when those resources can work together. The result is a new center within the enterprise, one around which more knowledge and activity are being organized.
In that sense, AI is becoming our zero milestone. More roads lead to it every day.
The difference is that this modern milestone does not simply tell us where the roads meet. It increasingly contains the value those roads were built to carry, which means the security of the center will matter to everything connected to it.
Rome built its roads because connection made the empire more powerful. We are building roads to AI for much the same reason.
We should learn from what happens when so much value converges in one place.