Self-improving runtime security firewall for AI agents that blocks prompt injection, tool misuse, and data exfiltration in real time.
Silmaril wraps AI agent inference calls with a low-latency firewall classifier that analyzes intent, application context, and execution state to block harmful outcomes before they materialize. Every discovered attack automatically becomes synthetic training data, updating defenses in under an hour and sharing protections anonymously across all deployments. The lightweight SDK wrapper works across major agent stacks including LangGraph and MCP clients, with managed or self-hosted deployment. The team has disclosed 15 critical vulnerabilities to OpenAI, Anthropic, Google, and Microsoft. Silmaril is a product of Silmaril.