Ethereum Foundation's AI Agent Discovers Protocol Code Bug During Testing

The Ethereum Foundation's AI (artificial intelligence) agent has revealed that it actually discovered a bug in the protocol code during testing.

Efforts to enhance the robustness of the Ethereum (Ethereum/ETH) blockchain, which has the largest amount of assets locked, are ongoing. The Ethereum Foundation's protocol security team recently conducted an experiment introducing an "AI agent" into the security testing of core software. While this initiative has garnered attention as a new approach to ensuring system safety, it has also highlighted the unique challenges posed by AI, particularly the burden placed on human reviewers.

Serious Bugs Discovered by Cooperative AI

In this test, multiple AI agents were run in parallel, each assigned specific roles. This configuration was inspired by research from the U.S. AI company Anthropic on "building C compilers using cooperative agent groups."

Specifically, the AI was divided into four roles for collaborative verification:

Scout Agent: Breaks down attack targets into specific verifiable hypotheses
Attack Agent: Attempts to track code paths and construct reproduction steps
Gap Completion Agent: Records progress and generates the next group of hypotheses
Verification Agent: Independently rechecks, eliminates duplicates, and assesses validity

As a result of this comprehensive exploration, the AI agents successfully discovered a serious vulnerability in the "gossipsub" layer of the P2P (peer-to-peer) library "libp2p," which is used for communication by Ethereum validators.

This bug could potentially cause a remote system crash of nodes, posing operational risks such as validators going offline and losing rewards if exploited. This defect was promptly fixed and published as "CVE-2026-34219," thus avoiding technical dangers.

The Surge of "Convincing False Positives"

While achieving the success of bug discovery, the protocol security team faced a significant bottleneck: processing the massive amount of "false positives (noise)" generated by the AI agents.

Traditional testing tools (like Fuzzer) output data when the system crashes, allowing humans to intuitively verify bugs. However, AI agents produce "plausible reports" with perfect structure that include attack scenarios, severity assessments, and proof-of-concept code. As a result, human engineers have to spend enormous time and effort distinguishing "fake bugs that look real at first glance."

According to the team's analysis, there are three main predictable patterns in the false positives generated by AI:

Tests Different from Production Environment: Crashes that occur only in debug builds with safety checks enabled, which do not affect actual users
Inaccessible Attack Paths: Reproduction code that manually inserts internal values that cannot be manipulated by attackers, as all external input paths are rejected
Empty Proofs of Formal Verification: While proving that the software operates correctly, these cases do not actually constrain the behavior of the target.

The Current State of AI in Web3 Security

Nikos Baxevanis from the Ethereum Foundation reported, "What surprised me was not the discovery of the bug itself, but how much effort was spent on the sorting work to distinguish between real and fake."

Moreover, while AI agents excel at inferring single moments of code, they still struggle to identify complex bugs, such as those that occur frequently in DeFi (decentralized finance) where "multiple normal steps are executed in a malicious order." The lesson from this test is that while AI security tools can become powerful assistants in accelerating the discovery of infrastructure bugs, human advanced judgment is essential for the final triage.

The introduction of AI has not taken away human jobs; rather, it can be said that the use of human time has shifted from hypothesis testing to building verification infrastructure and sorting information. Even in the modern era of evolving tools, the importance of strict triage as a human discipline remains unchanged.