Ethereum uses autonomous AI agents to improve blockchain security

The Ethereum Foundation shifts paradigm to automate cyber defense

The Ethereum Foundation is transforming its cyber security defense strategy. Its \"protocol security\" department has now deployed swarms of autonomous artificial intelligence agents to continue to attack its own networks. The goal is to discover, exploit and fix vulnerabilities before hackers do. The move, disclosed by the protocol security team, marks a major technological breakthrough-currently, the slightest flaw in a smart contract can lead to hundreds of millions of dollars in losses.

Brief introduction

The Ethereum Foundation deploys artificial intelligence agents to detect vulnerabilities before hackers do.
With these new tools, a critical network vulnerability has been discovered and fixed.
The AI agent group relies on a rigorous organizational structure to audit Ethereum\'s most sensitive infrastructure.
Researchers must now distinguish between real vulnerabilities and AI-generated false positives.

First tangible result in fighting cyber vulnerabilities

The preventive offensive launched by the Ethereum Foundation immediately proved its effectiveness, with a key vulnerability discovered at the core of the software on which the blockchain relies, and quantum resistance becoming a priority. The researchers confirmed that they planned a direct attack simulation against their own infrastructure, an attack method known as the \"red team test.\" In the official report, they shared preliminary findings, with the following highlights:

Targeting critical infrastructure: \"We launched coordinated AI agents for multiple system types that the network relies on, such as system software, cryptographic code, and contracts that must be airtight.\"
Discovery of real vulnerabilities: The scientists added unequivocally,\"These agents found vulnerabilities in production code that could be actually exploited.\"
Eliminate a major flaw: An exception was located in the \"Gossip sub libp2p\" protocol, the point-to-point network layer used by the Ethereum consensus client. This flaw can remotely trigger panic errors and threaten node stability. The vulnerability has been fixed and recorded on GitHub as the official reference number CVE-2026-34219.

In addition to simple testing, this experiment also revealed technological realities that human engineers had not expected. In fact, the use of large language models for software security has changed the nature of audit work itself, shifting the focus of work from crude research to key categories. Members of the Ethereum Foundation expressed surprise at the development: \"It\'s not surprising that agents can detect vulnerabilities.\" They pointed out,\"It\'s surprising that so little work is required to discover vulnerabilities, while distinguishing real vulnerabilities from seemingly real ones requires a lot of effort.\" This efficiency is in line with overall industry trends: In April last year, Anthropic\'s preliminary version of Claude Mythos model successfully identified 271 vulnerabilities in Mozilla\'s Firefox browser, demonstrating the computing power of these new tools.

Militarized organization of autonomous AI agent groups

To achieve such precision, the Ethereum Foundation has established a rigorous methodological architecture that allocates AI agents into highly specific role structures. The organization of these agent groups relies on four distinct and complementary functions: reconnaissance, vulnerability search, vulnerability filling, and final verification. One group of agents maps potential attack vectors, while another group strives to replicate the failure and directly tests the feasibility of the vulnerability against production code. The researchers emphasized the importance of this strict framework: \"There is a reason for this plan to exist.\" They argued that \"it forces specific and verifiable claims and clear definitions of the work done.\" An agent who has to compile observable evidence cannot just excuse him by saying \'it looks risk\'.\" This rigor eliminates the ambiguity common to traditional automated reporting.

Verification challenge against machine illusion

The increase in these detailed reports poses a major challenge for security teams, as the technical rhetoric of the machine does not guarantee its authenticity. Unlike traditional automated testing tools (called \"fuzzers\") that crash programs by simply injecting random data, AI agents write complex impact analyses and create proof-of-concept scenarios. The side effect is a large number of convincing false positives. To deal with this illusion, the foundation has established an absolute verification protocol. Researchers remind of an unchangeable golden rule: \"One rule is more important than all others.\" A candidate object can only be regarded as a discovery if there is an autonomous product that can reproduce the fault in real code and can be run by people who have not written the product.\" They concluded pragmatically: \"The regenerator does not read reports and does not care about the confidence level displayed by the model. It either works or it doesn\'t.\"

The transition to AI-assisted auditing heralds a new era for Web3. Recent history shows that this approach has borne fruit on a global scale. Last May, researcher Taylor Hornby used Claude Opus 4.8 to detect a key vulnerability in Zcash\'s Orchard privacy pool. The vulnerability, which was dormant for about four years, could have allowed fake ZEC tokens to be created without traces. By internalizing these technologies, the Ethereum Foundation has embraced a new operating paradigm. As its experts concluded: \"AI has not replaced security researchers, but shifted the focus of work.\" Leveraging these agent groups provides unprecedented code coverage, but in turn requires greater human insight. The researchers concluded: \"Agents allow us to cover much more than manual operations. In return, faced with a large number of confident assertions, they require more prudent judgment. As long as you remember that judgment is the real product, the process is worth it.\" Looking to the future, the resilience of blockchain will depend on humans \'ability to arbitrate machine diagnoses.