Silmaril

Silmaril is a self-improving firewall for AI applications that defends against prompt injection by continuously retraining on discovered attacks. It integrates with major agentic SDKs and operates as a real-time defense layer to block harmful outcomes before they materialize.

Silmaril wraps inference calls with a firewall classifier that analyzes user intent, application context, and execution state to block risky actions in real time. It uses autonomous threat hunting to discover attack chains, generates synthetic training data from those findings, and retrains the defense so updates deploy across deployments within under an hour. Integration requires five lines of code and supports managed or self-hosted controls across major agent stacks and node-level blocking.

Who it’s for: AI-native applications and platforms that deploy autonomous agents or toolchains and are concerned with prompt injection and related attacks (e.g., AI productivity apps, analytics platforms, code assistants).

Backed by Y Combinator; mentions of case studies, deployments, and ongoing retraining loops indicate traction and funding signals.