Stealthium Watches the GPU Layer Most Tools Miss

Security tooling has spent thirty years learning to watch the CPU. It reads process trees, syscalls, memory regions, and network flows, and it is genuinely good at it. Then the industry moved the most valuable computation in the enterprise onto accelerators, and almost none of that visibility came along. Stealthium, profiled by SecurityWeek on August 20, 2026, is a startup built specifically to close that gap.

• Stealthium describes itself as a security control plane for the GPU and accelerator layer of AI infrastructure
• It deploys an agent inside customer infrastructure and analyzes low-level GPU telemetry, kernel traces, and VRAM activity
• Target detections include cross-tenant leakage, model poisoning, and weight exfiltration on shared accelerator hardware
• The company is part of the NVIDIA Inception program and supports CUDA, PyTorch, TensorFlow, and Ollama workloads

What a GPU Blind Spot Actually Looks Like

Think about what runs on an accelerator in a modern AI deployment. Model weights sit in high-bandwidth video memory. Inference kernels execute in a scheduling domain that host-based endpoint tools do not instrument. Multiple tenants may share a physical device through partitioning. A conventional EDR agent watching the host sees a process holding a device handle and consuming power. It does not see what that process is doing inside VRAM.

That is the gap Stealthium is describing when it says GPUs are a blind spot in the modern AI stack. The company's framing, delivered by GTM advisor Chris Hosking, is blunt: "Our go-to thinking has always been that if you cannot see something happening, then nothing is happening. This is seriously dangerous, especially with accelerators."

Why Neo-Clouds Raise the Stakes

The risk concentrates in what the industry has started calling neo-clouds — AI-first specialized providers built around accelerators, distinct from the traditional hyperscalers. These environments are newer, they are multi-tenant, and they hold enormous concentrated value in the form of model weights that cost tens of millions of dollars to produce.

On shared accelerator hardware, the interesting attacks are the ones a host-level tool would miss entirely: reading residual data left in device memory by another tenant, subtly corrupting a training run so the resulting model behaves differently under specific inputs, or quietly copying weights out over the course of many small operations. Detecting any of those requires telemetry from the device itself.

How the Platform Works

Stealthium's approach is observability first, detection second. According to the company, the platform converts raw low-level GPU metrics and kernel traces into a higher-level abstraction it calls Hyperprints, then correlates security events against that runtime telemetry. Coverage spans single nodes, single clusters, multiple nodes, and multi-cluster deployments, which matters because a training job that spans dozens of GPUs produces behavior no single-device view can interpret.

The company advertises support across the NVIDIA software stack — drivers, toolkits, and CUDA — plus major AI frameworks including PyTorch, TensorFlow, and Ollama. It is sold on subscription and positioned as an extension of an existing security posture rather than a replacement for it, which is the correct posture for a layer this specialized.

The Pattern This Fits

Accelerator-layer security is becoming its own discipline, and Stealthium is one of several efforts pushing in that direction. NVIDIA shipped SkillSpector to scan AI agent skills for risk in August, the Open Secure AI Alliance brought more than forty firms together on shared defenses in July, and open-source projects like Hazmat are containing agents at the host level.

Stealthium is early — SecurityWeek's write-up does not disclose funding, founders, or named customers, and the company's own site does not publish performance figures. Treat the capability claims as vendor claims until independent evaluations appear. But the problem it names is real, and it is one that most security teams running GPU fleets cannot currently answer. More coverage in our AI security section.

Sources: SecurityWeek — August 20, 2026; Stealthium — August 2026.