Salus — AI Security Platform

Salus provides a runtime validation layer for AI agents, intercepting and checking every tool call before it executes. Part of YC W2026, the company was founded by Stanford CS roommates Vedant Singh and Kevin Pan, and has raised .7M in total funding.

The platform wraps around existing agent frameworks (OpenAI, Anthropic, LangChain, LangGraph, CrewAI) and validates proposed actions against an evidence cache of prior tool outputs and conversation history. Users define constraints using YAML, markdown, or plain English. When Salus blocks an incorrect action, it returns structured feedback enabling agents to self-correct — the company reports a 58% successful recovery rate on blocked actions.

Beyond runtime guardrails, Salus includes PII detection, content moderation, budget/loop protection, human-in-the-loop escalation, and an observability layer with real-time streaming of agent interactions. The evals system generates thousands of adversarial test scenarios to measure tool-call accuracy before deployment. Developers instrument their code with a simple Python decorator and define function dependencies.