Security Must Anchor the Enterprise AI Factory, Dell and Intel Argue

Enterprises are constructing artificial intelligence systems as production facilities for intelligence at scale. Yet the infrastructure powering these operations was not engineered with security as a foundational principle. Raw materials—energy, compute and data—flow through the system to yield intelligence in the form of tokens, but they also generate novel attack surfaces that conventional cybersecurity frameworks were never designed to address.

"AI factories introduce a new class of risks that extend beyond traditional cybersecurity models. In this world, data is not static, bounded or easily classified. Rather it is continuously generated, transformed and consumed across distributed environments," said Dave Vellante, chief analyst at theCUBE Research.

The systems underpinning these models operate according to deterministic logic but also function as adaptive, probabilistic engines that shift over time. Vellante noted that "Agents increasingly act autonomously, interacting with enterprise systems, executing workflows and making decisions with limited – or sometimes no – human intervention."

Agentic AI creates a new class of security risk

Dell Technologies Inc. and Intel Corp. are responding by embedding security into the infrastructure layer itself rather than bolting it on after systems go live. Both companies are working to establish data protection and AI security from the foundation upward.

"AI changes the whole game," said Steve Kenniston, senior cybersecurity evangelist for portfolio marketing at Dell. "There's the model inferencing; there's the training model, training data. There are the systems where people can do things like prompt injection; there's identity management that needs to be thought about. Every new application has a new attack surface."

The threat landscape extends beyond expanded attack surfaces. AI fundamentally alters what requires protection.

"AI brings additional threats," said Mukund Khatri, fellow and vice president of systems architecture at Dell. "The model … the LLMs look like code. They need to be protected like code, but they are essentially data. The model integrity is of paramount importance."

Autonomous AI agents are proliferating across corporate networks, raising urgent questions about governance as these systems gain the capacity to act without human oversight. Because humans create and manage these agents, organizations must evaluate how their autonomous actions and system access could expose sensitive data and critical infrastructure to compromise.

An incorrect response from a language model represents a problem; an incorrect action from an agent represents a more serious threat, according to Khatri. When systems operate independently, the potential for harmful errors or deliberate attacks multiplies, particularly because actions execute in real time with minimal opportunity for human intervention.

Identity becomes a critical control point for AI agents

Organizations are deploying AI at rapid speed, yet visibility into agent behavior and access permissions remains limited. Agent identity has emerged as a fundamental control layer; poorly managed agents could become entry points for adversaries. The Darktrace 2026 "State of AI Cybersecurity" report surveyed 1,540 cybersecurity leaders and practitioners across 14 countries and found that 92% expressed concern about AI agent security implications for their workforce, while 46% reported feeling unprepared to counter AI-powered attacks. The same report indicated that 87% believe AI is substantially amplifying malware sophistication and success rates.

Traditional security approaches centered on perimeter defense, static policy rules and human-paced response are proving inadequate in this environment. Techniques including model inversion, prompt injection and data poisoning can exploit weaknesses in models and data rather than conventional software flaws alone. Yet identifying where danger originates is only half the challenge. Speed is the other half.

Attackers can harness AI to discover vulnerabilities and construct exploits faster than before, compressing the window between vulnerability disclosure and active exploitation in ways enterprise security systems were not built to handle. AI models can scan code, pinpoint weaknesses and accelerate attack workflows. When agents traverse systems at high velocity, maintaining security becomes exponentially harder. Most security teams lack readiness for the complexity that AI workloads introduce.

Implementing least-privilege access will be essential to making agent deployment safer. Adapting governance practices such as lifecycle management and least-privilege access from human identities to agents does not require building from scratch because foundational frameworks already exist, according to Krista Case, principal analyst and practice lead for cyber resilience and security at theCUBE Research. The real challenge lies in scaling these approaches to accommodate a rapidly expanding population of non-human identities.

"As AI agents gain autonomy, identity becomes a critical control point," Case said. "Enterprises need to know which agents are operating, what they can access, what actions they can take and when those permissions should expire. Least privilege and lifecycle governance have to extend to agents as they become a growing class of enterprise identities."

This reality underscores why security teams must participate early as organizations move agentic systems into production. Governance frameworks and identity controls must be established before these systems begin processing sensitive information and interacting with business-critical workflows.

"I'm hearing more and more from our services organization that when they start to go into a customer environment and they start talking about AI and implementing AI, about 85% or 90% of those get stopped because the security team hadn't been involved up until that point," Kenniston said. "At Dell, we integrate security into everything that we do, from the supply chain through the chips to the device that gets delivered to you."

Dell constructs its AI infrastructure from inception with trust anchors embedded throughout its components, weaving cyber resilience into hardware and firmware rather than applying it as an afterthought. The company's Enterprise Hub leverages cryptographic image signing and SHA-384 hash verification to enable organizations to confirm the integrity and origin of AI model containers before they are deployed.

Securing the AI factory also demands attention at the network and operations layers. Dell's rack-scale infrastructure consolidates compute, networking and storage into a unified system where telemetry remains consistent and security teams gain visibility across the entire environment rather than monitoring isolated components. This matters because AI workloads generate continuous data flows between nodes, systems and models.

Visibility proves equally critical after an agent executes an action, especially when organizations must reconstruct what occurred and restore systems to a known secure state.

"As agents take on more responsibility for business processes, cyber resilience has to account for the state of the business," Case said. "Recovery will then require understanding what an agent did, what decisions led to that action and how to restore operations to a known good state."

Trusted hardware forms the foundation for AI security

Many enterprises are choosing to keep sensitive AI workloads on-premises, moving intelligence to the data rather than moving data to the cloud. This strategy grants organizations tighter control over sensitive information, infrastructure access and the location where AI processing happens. It also elevates the importance of securing the hardware substrate that supports models, data and inference operations.

"Security software and security measures running up the stack can't be trusted unless the hardware underneath them is trustworthy," said Mike Ferron-Jones, go-to-market lead for platform security and integrity at Intel. "The CPU is kind of the fundamental hardware root of trust in the entire security stack. Your choice of a CPU is your very first security decision that you are making."

Intel organizes its data center security capabilities into four categories: platform protection, confidential computing via SGX and TDX, software behavior enforcement through control flow technologies and encryption acceleration. These elements combine to establish a hardware foundation supporting the entire software stack.

Confidential AI environments isolate AI workloads inside a trusted execution environment featuring hardware-enforced isolation, cryptographic attestation and encryption keys managed by the organization.

Intel has created reference architecture work in collaboration with Nvidia Corp. that pairs CPU trusted execution environments, including Intel TDX, with Nvidia's confidential-computing capabilities for GPUs, extending confidential AI protections to GPU-accelerated workloads.

Post-quantum security enters the AI factory roadmap

Quantum computing progress is intensifying worry about the future viability of public-key cryptography systems in widespread use, though when a cryptographically relevant quantum computer will emerge remains unclear. A particular concern is the "harvest now, decrypt later" threat, where adversaries collect encrypted information today intending to decrypt it once quantum computers reach sufficient capability. Both Dell and Intel are preparing defenses against this scenario.

"By 2029, we expect that all cryptographic operations inside Intel platforms will be using quantum safe technology," said Ferron-Jones.

Intel's public roadmap separately targets full post-quantum cryptography compliance across all new platforms by 2030. Dell is advancing toward quantum-resistant protections and post-quantum cryptography across segments of its product range.

In an AI era where knowledge becomes tokens, security must underpin every technology layer.

"It must become the control plane that governs how intelligence is produced," Vellante said. "If you cannot secure the AI factory, you do not control the outcome."