SAN FRANCISCO – Ethereum Foundation developers have successfully used coordinated AI agents to identify a significant security flaw in the network’s messaging system, according to recent field notes from the Protocol Security team. While the experiment underscores AI’s potential in blockchain security, it also highlights a major hurdle: the immense amount of human effort required to distinguish genuine vulnerabilities from “convincing false positives”.
The Gossipsub Discovery
The vulnerability was found within gossipsub, a messaging protocol that allows Ethereum nodes to pass information to their neighbors. The flaw enabled a remote system to trigger an “impossible calculation,” causing the node’s software to shut down and taking the validator offline until an operator could restart it.
The Challenge of ‘Fluent Prose’
Despite the success, the Foundation warned that AI agents often produce misleading results with the same level of confidence as real ones. “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,” noted Nikos Baxevanis, the author of the report.
The team identified three recurring types of false positives:
- Test-only crashes: Issues that only exist in test builds with specific safety checks not present in the shipped software.
- Infeasible delivery routes: Attacks that require manual planting of dangerous values because external routes already reject them.
- Trivial proofs: Formal verification that is mathematically true but provides no useful information about actual software safety.
Sequential Exploit Limitations
The report also noted that AI agents struggle with “sequential transaction exploits”—vulnerabilities that span a series of individually valid steps. Recent high-profile drains, such as the Edel Finance exploit and the BONK governance attack, followed this pattern where each transaction appeared ordinary on its own.
To overcome this, the Ethereum Foundation has shifted its strategy. They now use AI agents primarily to propose suspicious sequences of actions, which are then rigorously tested and reviewed by human engineers to determine if they pose a real threat. This hybrid approach ensures that while AI handles the breadth of discovery, human judgment remains the final arbiter of truth.
Source: Coindesk
Editorial Note: This article was researched and drafted with AI assistance, then rigorously fact-checked, edited, and published by Miles. All content is strictly for informational and educational purposes only and does not constitute professional investment advice. Cryptocurrency and global financial markets experience severe volatility, sometimes swinging 50% or more in a single day. Invest only capital you can comfortably afford to lose, and always consult a certified financial advisor before committing funds. Read my full Disclaimer for more details.