AI agents are evolving from original assistants into autonomous "identities" that act on our behalf. Companies – including in Slovakia – are therefore building local AI data centers, but security here plays by new rules. A lecture by a Check Point expert showed why a firewall alone is no longer enough today and how to think about protecting agents, users, and applications.
AI agents: easily accessible, hard to predict
Practically anyone can program an agent today: inexpensive tools let you assemble workflows, connect memory, databases, send emails, or call payment gateways. The agent can connect to internal resources and RAG systems and has numerous input‑output integrations. It is precisely the broad permissions and connectivity that make one‑click convenience also a one‑click potential risk.
An agent is not a human and not even a deterministic machine: given the same stimulus it can respond differently and rely on different data. The result is the emergence of new risk scenarios, including so‑called "rogue" agents that act outside expectations. That is why many organizations are considering local AI data centers and stricter isolation of systems.
Regulation and today's security blind spots
According to the lecture, mandatory security testing of AI agents is approaching in the EU by around August 2026, which will also affect Slovakia. However, the pace of development is rapid and certainty is scarce; forecasts of full autonomy without "human‑in‑the‑loop" remain rather modest (an estimated minority share by 2028). What is certain is that companies will have to demonstrably test and document risks.
OWASP is already naming new classes of vulnerabilities specific to agents. Traditional tools – WAF, network and application protection, DLP or IAM – are hitting limits because the agent acts across layers. Security "guardrails" (guardrails) with tools like MCP are being extended in ways that cover only a small slice of risks and are easy to bypass.
How to protect AI "boxes": from perimeter to hardware
An AI server with a GPU may be physically small, but it behaves like an autonomous data center: it has databases, containers, applications, training models, and your data. It is therefore not enough to "put a firewall in front of it" or install an antivirus. What is needed is a combination of perimeter protection, microsegmentation, monitoring of ingress/egress and east‑west flows, and clear policies along three axes: the user and their data, the agent and data, and the applications themselves.
The lecture also emphasized the "AI vs. AI" approach and interventions at the hardware level. Integration into DPU cards (e.g., NVIDIA BlueField/DOCA) makes it possible to finely control the network, slow down privilege escalations, attempts to run malicious code or theft of models and training data, and to do so before the threat manifests higher up in the stack. Nevertheless, this is a rapidly changing field – what works today may change fundamentally in a few months.
