Aegis Protocol V5
An on-chain execution firewall for AI agents powered by Chainlink CRE.
What it is
Aegis Protocol V5 provides an on-chain execution firewall for autonomous AI agents. Giving agents direct custody of private keys introduces severe security risks. Aegis solves this by using Account Abstraction to strictly constrain agent permissions.
We utilize a Triple ERC stack to separate custody from execution. Funds are secured in an ERC-4337 Safe Smart Account. The AI agent operates using an ERC-7715 Session Key with defined allowances. All agent initiated transactions must pass through an ERC-7579 Executor Module before reaching the network.
When the agent attempts a trade, this firewall pauses execution and triggers the Chainlink Confidential Runtime Environment. Within a secure WASM enclave, we run parallel AI models (GPT-4o and Llama 3) to audit the destination smart contract for malicious logic like honeypots, hidden taxes, or privilege escalation.
If the multi model consensus flags the transaction as high risk, the transaction is rejected. The human owner can monitor these logs via our dashboard and manually revoke the agent's session keys if its risk profile becomes compromised.
How it Works
We built a custom ERC-7579 Executor Module integrated with an ERC-4337 Safe Account. Our decentralized AI firewall runs inside the Chainlink CRE. We wrote a WASM workflow that uses Confidential HTTP to securely fetch the destination contract data without leaking API keys, and it utilizes GPT-4o and Llama 3 in parallel for risk consensus.
We route all agent intents through Pimlico bundlers. Because Pimlico simulates the UserOperation off chain first, if our ERC-7579 firewall throws a TokenNotCleared error during the dry run, Pimlico drops the transaction entirely. The malicious trade never hits the Base Sepolia network, which prevents the user from paying wasted gas fees on failed trades.
Finally, a Next.js dashboard allows human capital allocators to manage agent budgets and monitor these security logs.
Links
Created by
- Vincent Beltrani