video1.0<iframe src="https://www.loom.com/embed/54181ca5021745dca97ba70cd6a7adbd" frameborder="0" width="1920" height="1440" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>14401920Loomhttps://www.loom.com14401920https://cdn.loom.com/sessions/thumbnails/54181ca5021745dca97ba70cd6a7adbd-2ba99d4f0af2a92b.gif206.4This Loom explains Agent Guard, a policy system that protects an autonomous treasury by preventing an AI agent from handling a master private key. The speaker highlights that a prompt injection could drain an entire wallet, and describes how Near protocol enables restricted key access via a new Create sandbox UD 25519 function call with strict congestion-level limits, daily spend cap checks, and contact whitelisting. They demonstrate scenario one on a Near testnet prep for compute credits, where the policy engine signs via restricted access keys and settles safely. In scenario two, an injected prompt to steal twenty NEAR from an unlisted contact is intercepted, cryptographic signing is blocked, and no NEAR is lost, with an on-chain emergency freeze switch and governance sliders for spend caps.