Security for agents that self-improves
This company is not in the game right now, so it has no score sheet.
Points come from real numbers anyone can check: new job posts, site traffic, software downloads and confirmed news. How scoring works
Facts from Y Combinator’s public list, last updated Nov 15. A free game. No real money. Points are a game score, not a judgment of what a company is worth.
Something wrong, or want this page gone? Fix or remove this page. Removal requests from a company's team are always honored. Founder of Silmaril? Claim this page for a Founder-verified badge and your own pitch.
Silmaril sells an AI firewall that learns from real attacks on a company's own AI agents, for the security teams who run them.
Think of it as Gmail's spam filter, but for the AI agents companies run.
AI agents read emails, web pages, and documents. Attackers hide instructions in that text to make agents leak data or take harmful actions. Security teams have no mature way to catch this, and hand-written rules fall behind new tricks.
The firewall sits between the agent and outside content, checking every input and action before it goes through. It studies attack attempts it sees in the customer's own setup and updates its detection on its own.
Security or platform teams at companies that deploy AI agents pay, usually through a yearly contract priced by usage volume or the number of agents.
AI security is a fast-growing category that is already crowded with startups and big security vendors. Silmaril competes by training continuously on each customer's own attacks, a narrower and more adaptive angle than generic filters.
Why now: Companies are moving from chatbots that only talk to agents that take actions, so a single successful attack now costs far more.
Written by AI from the company's public site. Not checked by the company. Something wrong? Fix it.