Stealthium and Tenstorrent Target the AI Infrastructure Security Blind Spot, Podcast

By Doug Green “If you can’t see something, you can’t secure it.” AI security discussions often focus on models, applications and data. But Chris Hosking, GTM Advisor at Stealthium, says organizations may be overlooking a critical layer: the accelerated computing infrastructure where AI actually runs. In this Telecom Reseller podcast, Hosking discusses Stealthium’s partnership with Tenstorrent and a broader emerging challenge for enterprises, cloud providers and operators — securing GPUs and other AI accelerators that traditional security tools often cannot fully observe. Stealthium describes itself as a runtime observability and security company for AI infrastructure. Its goal is to give organizations visibility directly into accelerated compute environments, below the point where many conventional security tools stop. “There’s an architectural gap here when it comes to controls,” Hosking says. Existing security instrumentation may provide extensive visibility at the endpoint, CPU and driver layers, but organizations increasingly need to understand what is happening inside the accelerator runtime itself. That becomes especially important as GPUs move from specialized computing resources to critical enterprise infrastructure. Hosking describes GPUs as environments that increasingly contain an organization’s “crown jewels” — including sensitive model weights, proprietary AI workloads and potentially valuable customer or corporate data. And GPUs are not simply black boxes performing calculations. They are programmable compute environments with memory, network access and the potential for persistence. That means they also create new attack surfaces. Stealthium’s recently announced partnership with Tenstorrent is designed to address that challenge by bringing runtime observability and security controls deeper into Tenstorrent’s AI infrastructure. Hosking points to Tenstorrent’s open, full-stack architecture as particularly important because it enables security and observability to be incorporated into the infrastructure rather than added later. The discussion also explores the security risks associated with shared GPU infrastructure. Hosking cites vulnerabilities such as Januscape as an example of how weaknesses involving virtualization and isolation can become particularly serious when multiple customers share physical accelerated-compute infrastructure. Customers need more than an architecture diagram saying workloads are isolated, he argues — they increasingly need evidence that the isolation is actually working. That leads to several basic questions AI infrastructure operators should be able to answer: What actually ran on the infrastructure? Was my workload isolated? Am I receiving all the compute capacity I paid for? Is another workload consuming resources? Is something behaving abnormally? Can I prove that the infrastructure is operating securely? Runtime visibility can therefore become more than a cybersecurity tool. It can also help organizations identify resource hijacking, inefficient GPU utilization and expensive compute capacity being consumed without operators realizing it. The issue takes on additional importance as enterprises pursue private and sovereign AI strategies. Organizations may own or control their infrastructure, but Hosking argues that sovereignty ultimately requires the ability to verify what is happening inside that infrastructure. Over the next six to twelve months, he expects security attention to move increasingly toward AI hardware and accelerated computing. As GPU adoption grows, vulnerabilities and attacks targeting these environments are likely to receive far greater scrutiny. For enterprises, service providers and cloud operators, the takeaway is straightforward: AI infrastructure is becoming too important — and too privileged — to remain a security blind spot. “If you can’t see something, you can’t secure it,” Hosking says. Learn more at Stealthium.io.