Keep pulling the thread on Stuart Audley.
Frontier AI models are not effective at finding security vulnerabilities in RTL and hardware design because they have not been trained on a sufficient amount of RTL data.
Caspia Technologies is developing two types of agentic workflows: an "agent-engineer loop" for human-in-the-loop assistance and an "agent-design loop" for full automation.
In a benchmark, Caspia's tools completed a full security verification workflow on the open-source Calyptra hardware root of trust RTL in less than one hour.
Stuart Audley estimates a manual security verification process for the Calyptra root of trust, which Caspia's tools completed in under an hour, would take an engineer weeks or even months.
Caspia Technologies offers a tool that performs security verification on RTL throughout the entire hardware design lifecycle.
Hardware security vulnerabilities like Spectre, Meltdown, fault injection attacks, and side channel attacks have proven that the assumption of hardware root of trusts being inherently secure is incorrect.
Caspia's technology uses AI to identify the location of sensitive assets like keys and configuration registers within a system design, which are then analyzed by deterministic tools to pinpoint vulnerabilities.
Caspia's verification tools can be integrated into a continuous integration (CI) flow, enabling security checks at every stage of the hardware design lifecycle.
The cost of finding and fixing a hardware security vulnerability at least doubles at each successive stage of the design lifecycle.
Daniel Nenny has observed from his experience that chip respins can lead to the failure of not just products, but entire companies.
Caspia's future-oriented agentic workflow aims to fully automate security verification for AI-generated RTL.
Caspia's Kodaks tool, an RTL static security analysis tool, uses two LLM-powered agents named Asset Assist and Report Assist.