The Protocol Security team from the Ethereum Foundation (EF) conducted a large-scale experiment: they deployed coordinated AI agents against critical components of blockchain infrastructure. System software, cryptographic code, and smart contracts were in the crosshairs. The results were unexpected even for the developers themselves.
The main conclusion drawn by the EF concerns not so much the number of bugs found, but the process of handling them. As the team noted, AI does indeed find bugs — this was not a surprise. What was surprising was how little effort was required to detect them and how much effort was needed to separate real vulnerabilities from false positives.
It is important to emphasize: artificial intelligence did not replace humans in the security audit process. It merely shifted the bottleneck. Previously, researchers spent most of their time searching for hypotheses; now, they spend it verifying a massive array of generated candidates. Triage, not generation, has become the new bottleneck.
System Architecture: Swarm Intelligence Instead of a Monolith
Protocol Security abandoned the classic scheme with one large AI agent managing the entire process. Instead, a network of specialized systems was deployed, working in parallel against a single repository. Some agents were responsible for initial code reconnaissance, others for finding potential vulnerabilities, filling gaps, and validation. Coordination was carried out through a shared repository and a version control system.
As explained by the EF, the time previously spent on forming and testing hypotheses is now redistributed to their mass evaluation: building an oracle, triage, maintaining a list of known issues, and disclosure. Agents excel at quickly scanning large code sections, tracing execution paths, and preparing materials for proof-of-concept. However, each candidate still requires independent reproduction on real code before it can be classified as a vulnerability.
Specific Findings and Limitations of the Method
A publicly disclosed example is the vulnerability CVE-2026-34219 in the Rust implementation of libp2p gossipsub. It is related to the handling of backoff expiry — the period during which a node temporarily restricts interaction with a peer. The error allowed remote process termination when processing a specially crafted PRUNE message with a near-maximum backoff value. The cause was unchecked arithmetic when adding Instant + Duration, leading to overflow.
However, the EF did not disclose the total number of real bugs found by the agents. The blog mentions several findings, but only one example is publicly named. Most generated candidates turned out to be false positives, duplicates, or issues outside the scope of the audit. The EF called this a normal part of the method, not a system failure. The goal is to quickly discard incorrect candidates and back up real ones with hard-to-dispute evidence.
Notably, AI agents perform worse with vulnerabilities that only manifest through a long chain of correct actions. Such logic requires not only finding a suspicious code fragment but also proving that the entire sequence of states is indeed reachable. This is a classic combinatorial explosion problem faced by all automated analyzers.
My professional opinion: The EF experiment is an important step in the evolution of blockchain security, but it clearly demonstrates that today AI is a powerful tool for generating hypotheses, not for verifying them. As long as the human factor remains a critical link in triage, full automation of security audits is a distant prospect. Investments in tools for rapid validation of candidates will be the next logical stage in the development of this technology.