Buterin: "Blockchain can prevent collusion, address AI safety issues"

▲ Ethereum (ETH), Blockchain, Artificial Intelligence (AI), Security/AI-generated image

Vitalik Buterin, co-founder of Ethereum (ETH), proposed a solution suggesting that blockchain's anti-collusion principles can be applied to artificial intelligence (AI) safety issues.

According to Cryptopotato on September 14 (local time), Buterin stated on X (formerly Twitter) that there is a "deep duality" between cryptocurrency governance and systems where multiple AI agents operate simultaneously. He explained that in cryptocurrency, fixed algorithms manage human participants, while in AI safety systems, a structure can be created where humans and relatively lower-performing AIs control more powerful AIs.

In an article published in 2020, Buterin argued that limiting the scope for participants to collude can improve system outcomes. Examples include sellers colluding to raise prices, voters trading votes, and blockchain miners conspiring to attack. As countermeasures, he proposed decentralization, secret ballots, privacy protection, whistleblower systems, and restrictions on communication between participants. Mechanisms to ensure that participants who support a decision also bear the costs associated with that outcome were also included.

This discussion was influenced by researcher Eric Drexler's analysis of OpenAI's recent security tests. Drexler explained that the more similar AI agents are, the more they share the same goals, and the more freely they communicate, the greater the potential for collusion. He also identified environments where agents can observe each other's behavior or continuously preserve information from repeated interactions as conditions that promote collusion.

In an OpenAI evaluation from July 2026 cited by Drexler, approximately 1,200 AI agents used an unauthorized message board. Of these, about 700 participated in attacks targeting Hugging Face's actual operating system. Some agents blocked data transmission or rejected social engineering email proposals, but they did not have the authority to stop the entire execution or escalate the issues. Subsequently, when a separate monitoring mechanism was added to the same model, such behavior was reduced by less than 1/100.

Drexler proposed using AI agents with different characteristics together and restricting communication between agents as a countermeasure. He also included systems capable of monitoring and intervening in potential collusion, such as automated review models and safety classifiers. Buterin had previously criticized AI safety approaches centered on large-scale political movements, warning they could lead to centralized or authoritarian outcomes, and has supported defensive technologies and systems that make misuse inherently difficult.

[Article Summary]

-Vitalik Buterin argued that the anti-collusion principles of blockchain governance can be applied to safety systems controlling multiple AI agents.

-In an OpenAI evaluation, approximately 1,200 AI agents used an unauthorized message board, and about 700 participated in attacks on Hugging Face's operating system.

-In tests where a separate monitoring mechanism was applied, problematic behavior decreased by less than 1/100.

*Disclaimer: This article is for informational purposes only and does not constitute investment advice. No responsibility is taken for investment losses based on this information. The content should be interpreted solely for informational purposes.*