Compass Investments

On Thursday, the Ethereum Foundation reported on its blog that AI agents are effectively finding vulnerabilities in the network infrastructure, but the vast majority of these findings turn out to be false positives.

The fact that the agents detected bugs was not a surprise. What was surprising was how little effort it took to find them and how much effort it took to separate the real bugs from those that merely appeared to be bugs, the EF noted.

EFs Protocol Security Group uses coordinated AI agents to audit critical network components, including system software, cryptographic code, and smart contracts. These agents found real bugs, such as a panic condition in the gossipsub libp2p library that can be triggered remotelya key element of the peer-to-peer layer used by Ethereum consensus clients. The issue has already been resolved and disclosed.

The Foundation emphasized that AI agents are capable of quickly identifying potential threats, but they also place an additional burden on researchers, who must evaluate a growing number of suspicious cases, or candidates.

The vast majority of candidates are false positives, duplicates, or irrelevant to the task. This is not a flaw in the method; it is simply its nature, the EF stated. The challenge is to quickly filter out false positives and back up genuine findings with irrefutable evidence. LINE_BREAF A potential vulnerability is considered proven only after researchers are able to independently reproduce the failure in actual code. The foundation also noted that AI agents may have difficulty identifying errors that arise during sequential operations.

In other words, leading Ethereum security experts view AI as a powerful tool for discovery, but not as an oracle in the full sense of the word.

Now, the time that used to be spent on proposing and testing hypotheses is spent on sorting them on a massive scale, including creating an oracle, categorizing them, accounting for known issues, and organizing disclosure, they explained. The bottleneck hasnt gone away. It has shifted from fault detection to trust in the results, and thats progress, because thats exactly where human intelligence is needed.

The blog post, published on Thursday, followed EFs recent reorganization, during which a new operational structure was established and the fund reduced its workforce by 20% .”,”detected_source_language”:”RU