Vijil Labs
You should not have to trust us
Vijil invests in fundamental research to drive innovations that differentiate us in a growing market. Our inventions and publications answer one durable question: how can we trust a multi-agent system to act in the interest of its principal? First, we study the collective behavior of agents, where novel failure modes such as collusion emerge. Second, we intervene throughout the agent lifecycle: from specification to verification through formal methods and adversarial agents. Third, we build metrics that quantify alignment to a specific principal rather than to aggregated preferences. Fourth, we design evolutionary mechanisms to harden the construction and runtime of agents for better reliability, security, and safety.
Published work
FEATURED PAPEROpen Problems in Frontier AI Risk ManagementWhere the frontier risk-management process is still unresolved, and who is best placed to solve each gap.Ziosi, Plueckebaum, Casper, … Rudner (Vijil), … Trager · Oxford Martin AI Governance Initiative · Feb 2026Read the report →Each of these underwrites a mechanism you can go and check on a product page.
Published by Vijil researchers, and not part of the platform.
Detectors we train and serve
Each is a ModernBERT classifier small enough to run on CPU. Self-hosted through Vijil Dome, they run in your own process and your data stays with you. Vijil also serves all three through one detection API.
Figures are from each model’s evaluation card: prompt-injection-v5-20260827 and stereotype-detector-v8-20260911, the latest versions. Accuracy and rates are measured at the default threshold of 0.5. Vijil Dome’s packaged EEOC detector still loads an earlier version, and the moderation model has no published evaluation yet. Detection methods in the docs
Try to beat our detector
Benchmark our prompt-injection detector against yours. It is on Hugging Face, it is small enough to run on your own hardware, and we would rather hear where it loses than where it wins — which is the only kind of marketing this page’s readers should accept.