The Ethereum Foundation said its Protocol Security team found and fixed a validator crash bug during an AI-assisted security review. In field notes released on July 9, researcher Nikos Baxevanis described a process that exposed both the usefulness and the limits of AI agents in protocol auditing.
Filtering believable false positives took most of the effort
Baxevanis said the hardest part was not discovering new bugs. It was deciding which findings were real. Traditional fuzzing tools usually return raw crash logs or stack traces, while AI agents produced detailed write-ups with call chains, severity assessments, and even working code samples. Some of those reports looked solid even when the underlying issue was not actually present.
According to the team, the most labor-intensive work was separating genuine vulnerabilities from outputs that only appeared credible. The Foundation grouped several recurring false positives into clear categories: crashes that happened only in test environments, attacks that worked only after manually inserted values, and formally valid proofs with little or no practical relevance.
AI agents still miss chained attack patterns
The team also found that AI agents struggled with vulnerabilities that emerge from a sequence of individually acceptable actions. That matters because recent high-value DeFi exploits have often relied on exactly this type of multi-step behavior. Broad coverage is useful. Accurate judgment on how separate actions combine is much harder.
For that reason, the Foundation now uses AI agents to suggest which scenarios deserve deeper testing, while human experts make the final call on whether a finding is a real vulnerability and whether it should be disclosed. The split is simple: the models expand the search, and researchers decide what holds up.
Foundation plans broader AI-assisted verification
The experiment is part of a wider shift inside the Ethereum Foundation toward AI-assisted verification, especially after significant staff reductions earlier this year. The Protocol Security team said AI can accelerate coverage, but human oversight remains necessary in evaluating and disclosing findings.
The article compared this approach with efforts at Anthropic and Cloudflare, which have also used AI agents to scale security research. In those cases as well, the final responsibility for interpreting results and taking action stayed with human analysts.
Ethereum launched in 2015 as an open-source blockchain for decentralized applications. The Ethereum Foundation, a nonprofit that supports Ethereum and related technologies, said the latest review shows AI now has a place in protocol security workflows, though the final gate on vulnerability confirmation still belongs to people.