AI infrastructure runs on GPUs, but most security tooling still watches the CPU. That creates a huge visibility gap. From the host, very different AI workloads can all look like the same trusted Python process using a GPU. As attackers target more of the GPU stack, CPU-centric tools like EDR, CSPM and SIEM can miss critical behaviour.
Closing that gap starts with understanding where existing controls lose visibility and what signals can restore it. Security teams need better telemetry across the AI infrastructure, along with sharper questions for GPU cloud and neocloud providers. They also need to account for unexpected attack paths, like model manipulation techniques that leave files on disk unchanged and generate no obvious network or driver activity.
In this session, AI & Accelerator Security Specialist Chris Hosking, will break down the GPU attack surface layer by layer, from orchestration and container runtime risks to driver exploits, side channels, cryptojacking, and ROME-class attacks. He will show what attackers gain, what current tools actually see, and what security teams can do to close the visibility gap.
Key takeaways:
– Understand where the GPU attack surface begins and how it expands across the AI stack.
– Spot the signs of ROME-class attacks and other threats that traditional controls are likely to miss.
– Pinpoint the gaps in CPU-focused security tooling and the telemetry needed to close them.
– Challenge GPU cloud and neocloud providers with the questions that matter before deploying sensitive workloads.
Ponentes: Chris Hosking, Director of Accelerator & AI Research | Stealthium