Salus, the Commit-Time Referee for AI Agents

Why it surfaced

Its central idea is unusually concrete: intercept the risky write, issue a decision receipt, and keep the agent in shadow mode before enforcement. A sample appointment update is stopped because three people share the same name, while the site reports controlled benchmark results showing lower prompt-injection success under runtime enforcement.

Salus checks an AI agent’s proposed action against policy and evidence before a backend executes it, then returns feedback for a safer retry.