The Ethereum Foundation has revealed that while AI agents can successfully identify software vulnerabilities, the biggest challenge in AI-assisted security research has become proving which reported vulnerabilities are genuine rather than finding potential bugs. According to the Foundation's Protocol Security team, recent experiments with coordinated AI agents uncovered real software flaws across systems that Ethereum depends on, but the organization said the majority of effort now goes into separating valid findings from convincing false positives. The team described these results in a technical post explaining how it has been testing AI agents against systems software, cryptographic libraries, and high-assurance smart contracts. The Ethereum Foundation emphasized that "agents finding bugs wasn't the surprise, the surprise was how little of the work went into finding them, and how much went into telling the real bugs from the ones that just looked real." The more instructive finding was not the bugs themselves, but the volume of noise that surrounded them - AI agents produced a large number of confident-sounding reports, and the majority turned out to be wrong, duplicated, or pointing to code paths that are unreachable in practice. As per the Foundation's latest post, "The core takeaway wasn't about finding bugs, it was about triage."
One confirmed discovery involved a remotely triggerable panic in the gossipsub component of libp2p, which forms part of the peer-to-peer networking layer used by Ethereum consensus clients. The Ethereum Foundation said the vulnerability was fixed and later disclosed as CVE-2026-34219 with credit to the team. The unexpected finding was not that AI could identify bugs, but that validating those reports consumed far more time than generating them. While agents can inspect source code, trace execution paths, and prepare proof-of-concept material, they also produce reports based on unreachable code, duplicate known issues, debug-only crashes, or weak formal proofs that fail to demonstrate a real security problem. The Foundation noted that "agents can read code, form hypotheses, trace call paths, and draft proof of concept artifacts. But it can also produce reports that look convincing while relying on unreachable code paths, debug only crashes, duplicate issues, or weak formal proofs that do not actually capture the intended property."
To reduce unreliable findings, the Ethereum Foundation said it deploys multiple AI agents against the same software repository, with each agent handling a different stage of the review process. Instead of relying on a central coordinator, the agents exchange information through the repository itself by sharing state in version control. The workflow begins with reconnaissance, where broad attack surfaces are narrowed into specific testable ideas, followed by hunting agents that follow each hypothesis through the code and attempt to build a working reproducer. Gap-filling agents track accepted and rejected reports to avoid repeating earlier work, while validation agents independently examine every candidate, remove duplicates, and determine whether a report qualifies as a legitimate vulnerability. The Ethereum Foundation described this approach as treating AI agents as "search tools rather than authorities" and emphasized that "the bottleneck moves from generating hypotheses to judging them through triage, known issue tracking, artifact validation, and disclosure."
At the center of the process, the Ethereum Foundation said one principle overrides everything else: a vulnerability does not count unless someone other than the reporting agent can reproduce it against the real codebase. According to the Foundation, this requirement removes reports built around impossible attack paths, debug-only failures, or formal verification results that appear mathematically correct without proving a meaningful security property. Every accepted report must identify a reachable target, define a clear security invariant, explain the failure mechanism, provide observable evidence, include a self-contained reproducer, and carry a deduplication key. Beyond technical validation, surviving candidates are also evaluated for practical exploitability, with the Foundation noting that AI agents remain inconsistent when judging exploit reachability and attack severity. The Ethereum Foundation emphasized that "the goal is to force every report into a concrete claim that can be tested against real code." Each surviving candidate is also checked for real world reachability and attacker cost, with the Foundation warning that "a bug any peer can trigger is different from one that requires special access or unrealistic resources."
The latest security update comes only weeks after the Ethereum Foundation completed a major internal restructuring. In a June 23 announcement, the organization said it had reduced its workforce by about 20%, with 54 employees leaving following a months-long review under its Mandate and Treasury Management Policy. According to the Foundation, the restructuring was intended to focus staff and resources on responsibilities that only the organization can perform while continuing long-term Ethereum development. The Ethereum Foundation warned that "agents are uneven. They can read specs, draft reproducers, and state invariants, but they struggle with reachability, severity, and bugs that unfold across valid sequences." For those complex scenarios, the team said agents work better as "guides for stateful test harnesses than as replacements." Separately, the Foundation has raised its maximum bug bounty from $250,000 to $1,000,000 for critical protocol vulnerabilities, with reports acknowledged within 48 hours and an initial assessment completed within one week. That expanded program signals how seriously the Foundation is treating protocol security as a strategic priority.
