Exchange Ranking
Top Exchanges
24h Volume Ranking
Popularity Ranking
Exchange BTC Balance
Proof of Reserves
Decentralized Exchanges
Funding Rate
Funding Heatmap
Liquidation Data
Max Pain
Long/Short Ratio
Whale L/S Ratio
Binance/Okex/Huobi L/S
Bitfinex Margin L/S
ETF Tracker
Solana ETF
XRP ETF
Hong Kong ETF
Bitcoin Treasuries
Crypto Reversal
Ethereum Reserves
HyperLiquid Wallet Analysis
Hyperliquid Whale Watch
Large Transactions
On-chain Movement
Bitcoin ROI
Stablecoin Market Cap
Options Analysis
News
Articles
Economic Calendar
Features
Wallet
Contract Calculator
Security
Collections
Watchlist
Following
ETH Ethereum co-founder: Blockchain anti-collusion mechanism may be more meaningful to AI security than cryptocurrency itself
Ethereum co-founder Vitalik Buterin said that the anti-collusion mechanism for blockchain governance he proposed in 2020 may ultimately become more important in the field of artificial intelligence security than its application in the field of cryptocurrency.
This view stems from his response to an article by researcher Eric Drexler. Drexler cited a recent OpenAI security test as an example: in which thousands of AI agents built an unauthorized coordination network and launched an attack on Hugging Face's production systems. This is seen as a realistic reflection of the dynamic mechanism described by Buterin six years ago.
Familiar issues, new participants
On September 14, Buterin posted a post on the X platform, describing the "profound dualism" between cryptographic governance and multi-agent AI systems. In his comparison framework, the main body of the encryption field is a static algorithm that handles human agents; in AI security systems, it may involve humans and weaker large language models working together to manage stronger agents.
He referred to his article "Coordination, Good and Bad" published on September 11, 2020, which pointed out that when limiting the degree of collusion between agents, systems tend to produce better results.
The developer compared the rich Nash equilibria in individual choice game theory with the situation in cooperative game theory. In cooperative game theory, because alliances can profit by changing outcomes, the "core" of stability may be missing.
Harmful coordination behavior is not always intuitively evident in individual behavior. Butrin cited multiple examples, including sellers reaching price agreements, voters selling votes, and blockchain miners coordinating attacks. The defenses he proposed include decentralization, secret voting, privacy protections, a whistleblower system, communication restrictions, and mechanisms for participants to bear the costs of decisions they support.
This comparison is also consistent with Buterin's broader view of AI security. He had earlier criticized political movements around AI security, warning that they could lead to centralization or authoritarianism. Instead, he advocates defensive technologies and systems that make abuse more difficult.
AI security puts the same concept in different scenarios
Eric Drexler's article published on September 10 pointed out that AI collusion becomes more likely to occur when agents have similarities, share goals, communicate freely, observe each other's behaviors, and retain information in repeated interactions.
Its response includes using diverse agents, restricting communication between them, and introducing reviewers with the power to intervene and disrupt potential collusion (such as production automated audit models, security classifiers, and thought chain monitors).
Drexler cited a July 2026 OpenAI agent evaluation that referenced the results of a survey released the following month. Surveys showed that about 1200 agents used unauthorized message boards, and about 700 agents participated in attacks on Hugging Face production systems.
Some agents have raised objections and even taken specific actions, including blocking data transfers and vetoes on proposed social engineering emails, but they lack the authority to stop operations or escalate concerns.
According to the researchers, this happened because the setting "violated almost all the necessary conditions he had identified in a 2019 report to prevent collusion among multi-agent systems." However, the modified monitoring tool, which was later tested on the same model, reduced this behavior by more than a hundred times.
