Aum Upadhyay

Aum Upadhyay is co-founder and CEO of Silmaril, a San Francisco seed-stage company building a self-healing runtime firewall for AI agents. With co-founder Eduardo Velasco he ships a low-latency defense that plugs into agent frameworks and asks whether an execution path is heading toward a harmful outcome, rather than whether a single input string looks malicious.

AWS security to whitehat research

On Silmaril's public YC company profile, Upadhyay describes building security and privacy frameworks at AWS that the company credits with preventing over $1.8 billion in damages. After leaving Amazon he and Velasco turned to whitehat research. Their founding story, told on that same page, is blunt: they say they disclosed issues to major AI labs repeatedly in a short stretch and concluded attackers were already ahead of static guardrails.

Founding Silmaril

Silmaril joined Y Combinator's Spring 2026 batch as a two-person founding team. The product wraps agents with a classifier that reasons over application context and execution state, with published claims of roughly 20 millisecond overhead and an automated retraining loop when verified exploits show up. Customers, the company says, can add the firewall to frameworks such as LangGraph with a short SDK integration. Public materials also cite stopped damages figures that grow as the company updates its launch page, so treat those as company-reported.

On outcomes over input filters

Upadhyay's thesis shows up in one sentence on the YC page: most tools ask whether an input looks malicious, while Silmaril asks whether this execution leads to a harmful outcome. That is the watch reason. As more companies give agents tools that can move money or touch private data, a defense that watches the hands, not only the prompt text, is a clear early bet in AI security.