One Wrong Tool Call Can Cause Irreversible Damage

As companies deploy AI agents that touch real production systems — moving money, editing records, hitting customer-facing APIs — a single wrong tool call can cause irreversible damage. Existing safety approaches are either brittle prompt-level instructions or blunt blocks that kill the task entirely, leaving a gap in reliable, enforceable controls at the exact moment an agent acts.

What They're Building

Salus is a policy-aware runtime proxy that sits between an AI agent and its tools, making commit-time the moment of judgment. It intercepts every tool call before execution, validates it against versioned (YAML or natural-language) policies and a contextual evidence cache, and — rather than simply blocking — clarifies, repairs, rewrites, or escalates the action. When an action is blocked, the agent receives structured feedback to guide a corrected retry. The system also includes PII detection and budget-protection safeguards.

Growth

Salus was founded by Vedant Singh and Kevin Pan as part of Y Combinator's Winter 2026 batch, raising $4M. Early benchmark traction includes a reported 58% recovery rate on blocked actions (the agent completes the task correctly after structured feedback), a 60% cost reduction on τ²-bench, and a 52% misalignment reduction on ODCV-Bench across 12 models.