Ethereum Foundation Uses AI Agents to Identify Real Vulnerability in Protocol

The Ethereum Foundation recently conducted an experiment where AI agents were used to test the protocol's security. The goal was not to replace human researchers but to provide them with a larger pile of evidence to examine.

The experiment produced several convincing reports, including one that identified a real vulnerability in libp2p's Gossipsub component. This issue could cause a remotely triggered panic and potentially take an affected process offline.

A human researcher had to reproduce each claim, study the reachable code path, remove duplicates, and separate genuine risks from confident-sounding errors before it was confirmed as a finding.