Loops HouseAruna

NEAR HACKATHON
Team
1 member- MUOwner
Muhammad Aris
Overview
Aruna is a risk layer for AI agents that spend money on NEAR. It tells an agent whether the contracts it pays can still change, keeps the agent inside limits it cannot break, and pays out when a covered contract's code changes anyway.
The problem: agents pay contracts thousands of times with no human in between, and a contract they were approved to use can change underneath them. Across the 1,000 most-called mainnet contracts, 44.2% of calls went to contracts whose code a full-access key holder can replace at any moment, and only 0.4% to a contract with no detected way to change. On NEAR it can also happen invisibly: an account can run code published by another account, and when that code is replaced, no transaction touches the account and the hash everyone monitors (view_account.code_hash) never changes. Three of those 1,000 work this way today.
For companies that run agents for clients:
- Check: any agent or developer can ask whether a contract can still change, from a web page or a keyless tool.
- Contain: one vault per client. The client owns the funds, the company operates the agents, the vault holds zero full-access keys, and each agent is a function-call key anyone can inspect in one public call. When we prompt-injected an agent on testnet, the vault refused it five times and paused it on its own.
- Cover: underwriter capital pays out when a covered contract changes. On testnet a contract was changed, reported 31 blocks later, and a claim paid 0.08 NEAR. Today one pool backs every risk class; next, underwriters choose the risk they carry.
How NEAR fits: function-call keys are the containment primitive, and global contracts are both the blind spot we measure and how our vaults are deployed with no full-access key.
Live on NEAR testnet, open source; each README step says what to run and what you should see. Nothing holds funds on mainnet yet, and two trust assumptions remain: an attestor and an arbiter. An audit and first pilots come next.