The Protocol Security team, part of the Ethereum Foundation (EF), conducted a large-scale experiment using coordinated AI agents to find vulnerabilities in critical components of the blockchain infrastructure. The targets of the audit included system software, cryptographic code, and smart contracts.
The main conclusion drawn by the researchers was non-obvious: the key problem was not in detecting bugs, but in their triage. As noted by the EF, the agents do find vulnerabilities, but the lion's share of time is spent not on searching, but on separating real threats from false positives.
Interestingly, the team abandoned a monolithic approach with a single large AI agent. Instead, they deployed a system of several specialized agents working in parallel against a single repository. Some were responsible for initial code reconnaissance, others for identifying potential vulnerabilities, filling gaps, and validation. Coordination was carried out through a shared repository with state exchange via a version control system.
According to the developers, the agents excel at generating search directions: they can quickly scan large volumes of code, trace execution paths, and prepare materials for proof-of-concept. However, each candidate requires independent reproduction on real code before it can be classified as a vulnerability.
Practical Results and Limitations
A publicly disclosed example of the AI agents' work was the vulnerability CVE-2026-34219 in the Rust implementation of the libp2p gossipsub protocol. It was related to the handling of backoff expiry — a period of temporary communication restriction between nodes after certain network messages. The error allowed remotely triggering a process crash via a specially crafted PRUNE message with a near-maximum backoff value, leading to an overflow due to unchecked arithmetic when adding Instant + Duration.
The EF did not disclose the total number of real bugs found, only mentioning a few findings. Meanwhile, most generated candidates turned out to be false positives, duplicates, or issues outside the scope of the audit. The foundation considers this a normal part of the methodology, not a system failure.
Notably, AI agents show weak results when dealing 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 actually reachable.
The EF publication coincided with a major reorganization of the foundation. In June, the staff was reduced by 20%, and the budget was cut by approximately 40%. Vitalik Buterin confirmed the transition to a long-term capital management model. Additionally, the Protocol Support team, responsible for coordinating protocol development, supporting EIPs, and educational initiatives, was disbanded.
Against this backdrop, former EF employee Trent Van Epps warned of the risk of a "slowly escalating funding crisis" over the next three to nine months.
Analyst's Comment: The EF experiment demonstrates a mature approach to integrating AI into blockchain security. Instead of hype around "autonomous bug hunters," we see pragmatic integration: AI takes over routine reconnaissance, but the final decision remains with humans. This is the right direction, especially given that false positives and complex multi-step attacks remain the Achilles' heel of the agents. However, against the backdrop of budget cuts and EF reorganization, the question arises: could this technology become an attempt to compensate for a shortage of human resources?