The triage is the product: running AI agents against Ethereum's protocol code

The Ethereum Foundation’s Protocol Security team coordinated AI agents to audit core systems like Ethereum’s consensus clients, cryptographic code, and contracts. The agents identified real vulnerabilities, including a remotely-triggerable panic in libp2p’s gossipsub, a peer-to-peer layer used by consensus clients, which was fixed and disclosed as CVE-2026-34219 with credit to the team. The focus was not on discovery alone but on distinguishing genuine bugs from false positives, a challenge that dominated the effort.

The team structured the process around a repeatable pipeline: parallel agent searches, shared coordination via version control, independent validation, and strict triage. The method mirrors approaches used by Cloudflare and Anthropic, emphasizing reproducibility over tooling specifics. A key rule required each candidate bug to include a self-contained reproducer that runs against the real code, ensuring findings were verifiable and reducing reliance on model confidence.

Most candidate reports were incorrect, duplicates, or out of scope, highlighting the importance of automated triage. Each surviving candidate underwent two independent checks: assessing exploitability in normal configurations and comparing attack cost to potential network impact. The acceptance rate varied by codebase maturity, with heavily audited code yielding fewer valid findings, while less-explored or formally verified code produced more candidates.

The team found that the hardest part was not generating candidates but validating them. Agents excelled at one-shot reasoning but struggled with multi-step bugs requiring sequential validation. The agents’ role was best suited to suggesting sequences for stateful testing, not replacing it entirely. The approach underscored that AI audits are most effective when paired with rigorous, automated triage and expert review.