Silmaril Security, a portfolio company in Robinhood Ventures Fund II, is building runtime security for self-improving AI systems, the class of software that modifies its own code and behavior as it operates. The fund lists the startup among its AI security holdings ahead of the Aug 13 NYSE listing of RVII, placing it in a cohort that also spans digital identity and counter-drone defense. The company’s own positioning puts the mission in execution terms: protect the runtime before an agent’s drift turns into an incident.
The security case is straightforward. A model that can rewrite its own logic can outrun signature-based defenses, static sandboxes, and behavior monitoring built for binaries that never change. Because the threats are behavioral rather than signature-shaped, detection has to reason about drift and intent instead of known-bad hashes. Silmaril’s approach is to watch the execution environment in real time, detect when an agent begins acting outside its guardrails, and intervene before a compromised model can reach other systems. That is the same problem security teams face with connected tools and agents that hold credentials, only faster.
Runtime protection of this kind is early enough that most enterprises have no dedicated tooling for it, which is part of why the fund can enter at venture pricing. The bet mirrors a broader shift as AI agents gain access to production systems, sensitive data, and payment rails: the perimeter stops being a network boundary and becomes the behavior of the model itself. The fund’s prospectus describes the holding simply as runtime security for self-improving AI, and the category has no established vendor roster yet.
For security teams the takeaway is timing. Self-improving runtimes are coming regardless, and the vendors that ship working runtime controls before the first wave of agent compromises will define the category. Silmaril, backed by RVII, is positioning to be one of them.
