My team, Protocol Security at the Ethereum Foundation, conducted a large-scale experiment: we deployed coordinated AI agents against key components of blockchain infrastructure. The targets included system software, cryptographic code, and smart contracts. The results were unexpected — and highly revealing.
Rethinking the Role of AI in Security
The main takeaway is not that the agents found bugs. That was expected. What surprised us was how little effort was needed to find them and how much was required to separate real vulnerabilities from false positives. AI did not replace humans; it shifted the bottleneck. Previously, researchers spent a lot of time generating hypotheses; now, they spend it on verifying a massive stream of candidates.
We abandoned the single "super-agent" approach. Instead, we deployed several specialized systems working in parallel against a single repository. Some handled code reconnaissance, others focused on vulnerability discovery, filling gaps, and validation. Coordination occurred through a shared repository and version control system. This architecture dramatically accelerated the process but did not eliminate the need for manual review: each candidate had to be independently reproduced on real code.
What Was Found
A publicly disclosed example is the vulnerability CVE-2026-34219 in the Rust implementation of libp2p gossipsub. It relates to the handling of backoff expiry — the period during which a node temporarily restricts interaction with a peer. The error allowed remote process termination via a specially crafted PRUNE message with a near-maximum backoff value. The cause was unchecked arithmetic when adding Instant + Duration, leading to overflow.
However, we are not disclosing the exact number of real bugs found by the agents. The blog mentions several findings, but only one is publicly named. Most candidates turned out to be false positives, duplicates, or issues outside the scope of review. This is not a failure but a normal part of the method. The goal is to quickly discard incorrect ones and reinforce real ones with evidence.
Limitations
AI agents struggle more with vulnerabilities that only manifest through a long chain of correct actions. Such logic requires not just finding a suspicious code fragment but proving that the entire sequence of states is actually reachable. This remains a human responsibility.
The publication comes amid a foundation reorganization: a 20% staff reduction, updated governance structure, and a 40% budget cut with a shift to a long-term capital management model. The Protocol Support team, which coordinated protocol development, has also been disbanded. In March, we solidified the principle of minimal intervention in network development, and by 2029, we plan to implement quantum protection.
My comment: The experiment confirms that AI is a powerful tool for automating routine tasks in security auditing, but it cannot replace deep analytical thinking. The key challenge is not finding bugs but filtering out noise. While agents generate a mass of candidates, the real value remains with humans capable of separating the wheat from the chaff.