Key facts
- The Ethereum Foundation is using AI agents to conduct security audits on its network.
- These AI agents are designed to act as 'red teams,' attacking systems to find weaknesses.
- A peer-to-peer software vulnerability in libp2p's gossipsub was discovered and fixed.
- AI agents are organized into specialized roles for reconnaissance, hunting, gap-filling, and validation.
- The process requires human researchers to filter AI-generated findings for accuracy.
The Ethereum Foundation has initiated a novel approach to network security by deploying swarms of AI agents to proactively hunt for vulnerabilities. Researchers on the Protocol Security team are using these agents to attack critical systems, including cryptographic code, protocol logic, and smart contracts, in a process known as red teaming.
This AI-driven security initiative has already yielded results, with the discovery of a remotely triggered panic in libp2p's gossipsub, a component of Ethereum's peer-to-peer layer. This vulnerability was subsequently fixed and publicly disclosed as CVE-2026-34219.
The foundation's researchers noted that while AI agents can efficiently scan codebases and identify potential exploits, a significant part of the process involves human researchers filtering the AI's findings to distinguish genuine bugs from false positives. The AI agents are structured with specialized roles, such as reconnaissance, hunting, gap-filling, and validation, to systematically probe for weaknesses.
This development aligns with a broader trend of AI integration in cybersecurity. Previously, AI tools, such as Anthropic's Claude, have assisted in uncovering numerous vulnerabilities in other projects, including a critical flaw in Zcash's privacy pool identified by researcher Taylor Hornby.