Salus is a runtime guardrails platform from YC W26 that validates AI agent actions before they execute, preventing harmful, unauthorized, or policy-violating behaviors in production systems. Teams define policies using YAML, markdown, or plain English descriptions, and Salus intercepts agent actions in real time to verify compliance. The platform uses evidence grounding to check whether agent decisions are supported by factual context, catching hallucination-driven actions that would otherwise reach production systems.
A key differentiator is Salus's structured feedback mechanism that provides agents with specific guidance when an action is blocked. Rather than simply rejecting and halting the workflow, Salus tells the agent what went wrong and how to correct it, achieving a 58% recovery rate where blocked agents successfully self-correct and complete their tasks. This approach reduces the cost of policy compliance by up to 60% compared to naive filtering approaches, while decreasing misalignment by 52% on frontier models including GPT-4 and Claude.
Founded by Stanford CS alumni Kevin Pan and Vedant Singh, Salus addresses the growing need for production-grade safety infrastructure as enterprises deploy AI agents with access to sensitive tools, data, and external services. The platform includes built-in protections for common risk vectors including PII detection to prevent data leakage, budget controls to cap agent spending, and human-in-the-loop escalation for high-stakes decisions. Salus positions itself as the safety layer between AI agents and the real-world actions they perform.