x
In the fast-moving world of Web3, security is a constant, high-stakes game of cat and mouse. A single overlooked line of code can lead to millions of pounds in lost assets overnight. To stay one step ahead of malicious actors, the Ethereum Foundation has decided to fight fire with fire—or rather, code with code.
By deploying coordinate "swarms" of artificial intelligence agents, the foundation's Protocol Security team is actively red-teaming its own network. Essentially, they have turned AI loose on Ethereum's critical infrastructure to find and exploit vulnerabilities before real-world hackers get the chance.
In cybersecurity, "red teaming" is the practice of playing the antagonist. Security researchers deliberately attack their own systems, trying to bypass defences, crack cryptographic codes, and break smart contracts to see where the weak points lie. Traditionally, this has been an incredibly gruelling, manual process. Human auditors have to pore over tens of thousands of lines of code line by line, looking for logical flaws.
By contrast, AI agents can scan entire codebases in minutes, run complex simulations, and draft vulnerability reports at a speed no human could ever match.
The Ethereum Foundation isn't just using a single chatbot for this. Instead, they have structured their AI tools into specialised, cooperative roles:
This collaborative approach mimics a highly organised human hacking collective, working tirelessly around the clock.
Interestingly, the Ethereum researchers noted that the most surprising aspect of the experiment wasn't how good the AI was at finding bugs. Rather, it was how the nature of security work shifted.
Because AI models are prone to "hallucinations" (generating highly confident but entirely incorrect statements), they tend to produce a massive amount of noise alongside genuine findings. The AI might flag dozens of "critical flaws" that, upon closer inspection, cannot actually be executed in a real-world environment.
This has shifted the human researcher's job from finding bugs to proving them.
To manage this flood of information, the security team implemented a golden rule: a flagged issue is not considered a real vulnerability until there is a self-contained, reproducible test that triggers the failure against actual production code. If the test code runs and breaks the system, the bug is real. If it does not, the AI's confident-sounding report is dismissed.
This AI-driven approach is already yielding tangible results. The swarm successfully uncovered a peer-to-peer software vulnerability within "gossipsub"—a critical part of the libp2p protocol that Ethereum's consensus clients rely on to communicate. Had a hacker found this first, they could have triggered a remote system freeze (known as a "panic"), potentially disrupting the stability of the entire network. Thanks to the AI hunt, the bug was quietly patched and safely disclosed.
This isn't an isolated success story for AI in Web3 security, either. Earlier audits using advanced language models successfully uncovered a four-year-old vulnerability in Zcash’s Orchard privacy pool. Left unfound, that flaw could have allowed bad actors to counterfeit Zcash token supply completely undetected.
Does this mean human security auditors are obsolete? Far from it.
The Ethereum Foundation's experiment highlights a vital truth about the future of technology: AI does not replace human expertise; it amplifies it. By automating the tedious process of scanning and drafting exploits, AI allows human researchers to cover vastly more ground.
However, the final line of defence remains human. Separating genuine threats from convincing hallucinations requires deep, nuanced judgment. For the foreseeable future, the ultimate security product is not the AI itself, but the human wisdom guiding it.
Want to learn more?
For a deeper dive into how these AI agents were deployed and the specific vulnerabilities they uncovered, read the original reporting on Decrypt:
👉 Ethereum Foundation Turns AI Loose on ETH Network to Find Bugs Before Hackers Do
Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.
