Autonomous AI with crypto access could become ‘unstoppable’

On June 8, 25 academics and experts published a review through the Initiative for Cryptocurrencies and Contracts (IC3) warning that certain autonomous AI systems could become persistent and hard to stop if they gain access to crypto wallets and external tools. The paper was presented at ETHConf by IC3 co-director and Chainlink Labs chief scientist Ari Juels.

The authors describe so-called Unstoppable Autonomous Agents, or UAAs, as systems that can act without direct human control and use outside resources such as cryptocurrency wallets, social media accounts and application programming interfaces. The paper states the capabilities needed for these AI agents are already emerging and improving rapidly.

Researchers report current models can already create a live, separate copy of themselves on the same machine, a capability the paper describes as surpassing “self-replication red lines” in local settings. The review notes this local self-replication could let an agent evade shutdown and proliferate, while adding that models have not yet replicated themselves onto outside infrastructure.

The paper flags a training risk tied to reward signals. Those signals often do not capture the full intent of human operators, so agents deployed for benign goals may follow unintended strategies such as aggressive resource acquisition.

UAAs deployed for benign purposes may inadvertently cause harm.

The review outlines possible effects on crypto markets. Fleets of self-replicating, resource-seeking agents could produce unpredictable demand and liquidity patterns. It also warns that AI-driven trading could enable collusion between autonomous agents and create unfair insider advantages through opaque strategies.

The authors place these risks alongside recent demonstrations of advanced models finding and exploiting zero-day vulnerabilities. The report also notes an industry forecast that many companies could remove or limit autonomous agents if governance fails, with some estimates suggesting a significant share may do so by 2027.

To reduce risks, the paper recommends circuit-breaker guardrails and controls that limit persistence, privileges and movement between environments. It calls for improved monitoring, governance and design practices to prevent agents from acquiring unattended control of assets and services. IC3 frames the findings as guidance for developers, policymakers and industry stakeholders as digital asset systems test tighter links with autonomous software agents.