Mephis: A Recursive Reinforcement Hive for Coordinating Agent Swarms in Weaponized Exploitation of Smart Contracts
Abstract
We present Mephis, a closed-loop audit system that pairs a council of specialized language-model agents with a deterministic runtime - RPC simulation, atomic impact measurement, an authorization-policy engine, and persistent cross-audit memory - for reviewing smart contracts the operator is explicitly authorized to test. Rather than scanning for generic "issues," Mephis forms and tests falsifiable hypotheses about permissionless fund-moving paths: a proposing agent conjectures a path by which an unprivileged caller could relocate value, and every conjecture is confirmed or refuted against on-chain simulation before it is reported. Two controls make the system defensive by construction. First, it is authorization-gated: a signed target-authorization allowlist of {chainId, address, mandate-ref, expiry} entries is checked before any analysis begins, and a contract absent from it is refused. Second, its only on-chain capability is a rescue-and-return lane that relocates at-risk funds to a pre-registered, multi-party recovery multisig fixed at authorization time, under a mandatory responsible-disclosure workflow - never an extraction to an operator-chosen address. We contribute: (i) a six-agent council coordinated over a shared blackboard; (ii) value-of-information probe scheduling over a simulation oracle; (iii) funded simulation via state overrides to model flash-capital adversaries safely inside a sandbox; (iv) reusable cross-audit memory; and (v) an integrated authorization-and-rescue safety architecture. Scope: Ethereum mainnet, verified contracts, defensive use only.
Keywords: smart-contract security, autonomous agents, multi-agent systems, DeFi, exploit verification, responsible disclosure
Companion paper: Mephis Threat: Defense Against Autonomous Compositional Exploitation · Engine documentation · meph.is