The Ethereum Foundation has deployed a series of AI agents to attack Ethereum software to find vulnerabilities before attackers, described as AI-driven vulnerability discovery on Ethereum. The agents have produced actionable findings, including one remotely triggered panic that was fixed and disclosed on GitHub as CVE-2026-34219, and the Foundation requires a self-contained reproducer before a candidate is treated as a confirmed finding. As a related benchmark, Anthropic’s Claude Mythos reportedly discovered 271 vulnerabilities in Mozilla’s Firefox in April; the article is dated July 9, 2026.
The Ethereum Foundation’s AI agents are organized into distinct roles: reconnaissance, hunting, gap-filling, and validation. Some agents search for attack paths across systems while others focus on reproducing failures and verifying issues against production code. The agents target cryptographic systems, protocol code, and smart contracts used by the network. Those agents can scan entire codebases, test potential exploits, and generate findings for human review.
Generated findings are subject to review because AI-produced reports can be convincing even when they are incorrect, requiring filtering of duplicates, false positives, and non-exploitable issues. The agents can produce vulnerability reports, assess impact, and create proof-of-concept tests, capabilities distinguished from traditional fuzzers. A procedural rule requires a self-contained artifact that reproduces a failure against the real code before a candidate is treated as a confirmed finding. The AI program has produced actionable results, including a remotely triggered panic in libp2p’s gossipsub that was fixed and disclosed as CVE-2026-34219.
Red teaming is described as security researchers attacking a system to uncover weaknesses before malicious hackers. The Ethereum Foundation applied this concept by deploying AI agents to attack Ethereum software to find vulnerabilities before attackers. The AI agents are organized into roles—reconnaissance, hunting, gap-filling, and validation—and some search for attack paths while others reproduce failures and verify against production code. The agents target cryptographic systems, protocol code, and smart contracts used by the network.
AI agents can scan entire codebases, test potential exploits, and generate findings for review. The program can produce vulnerability reports, assess impact, and create proof-of-concept tests, capabilities that the article contrasts with traditional fuzzers. AI-generated findings can be convincing even when they are wrong, requiring filtering of duplicates, false positives, and non-exploitable issues. The Foundation enforces a rule that a candidate is not a confirmed finding until there is a self-contained artifact that reproduces the failure against the real code and runs for someone who did not write it.
The Ethereum Foundation deployed AI agents to attack Ethereum software to find vulnerabilities before attackers, and those agents discovered a remotely triggered panic in libp2p’s gossipsub. That remotely triggered panic is presented as a specific example of successful AI-driven discovery on Ethereum-related code. The discovery occurred as part of a program that targets protocol code, cryptographic systems, and smart contracts.
The vulnerability was fixed and publicly disclosed on GitHub under the identifier CVE-2026-34219. The Foundation enforces a rule that a candidate is not a confirmed finding until there is a self-contained artifact that reproduces the failure against the real code and runs for someone who did not write it. The article cites this issue and its correction as an actionable result produced by the AI agents.
Validation in the Ethereum Foundation’s AI-driven vulnerability discovery requires AI-generated candidates to be converted into reproducible artifacts before being treated as confirmed findings. Agents produce vulnerability reports, assess impact, and create proof-of-concept tests, and some agents are specifically tasked with reproducing failures and verifying issues against production code. Because AI-generated findings can be convincing even when incorrect, the process includes filtering duplicates, false positives, and non-exploitable issues prior to confirmation. The schema enforces a definition of done where a self-contained reproducer must run the failure against the real code for someone who did not write it.
The article states that a candidate is not a finding until a self-contained artifact reproduces the failure and runs for someone who did not write it. Together with agents’ abilities to scan entire codebases, test potential exploits, and produce proof-of-concept tests, this validation rule is presented as central to the Ethereum Foundation’s AI-driven vulnerability discovery on Ethereum.
