Advances in artificial intelligence (AI) are transforming software development. A report by Argentinian company Lambda Class points out that this change creates opportunities and risks for the cryptocurrency ecosystem, especially when automated systems interact directly with real money without continuous human intervention.
In a document published on January 23, the company focuses on developing tools for Ethereum and proposes the use of AI agents to manipulate the cryptocurrency. Introducing new vectors of security flaws. These are factors that were not considered in the original design of the infrastructure.
According to the report, the introduction of AI agents (programs that can autonomously make decisions and perform actions) changes key assumptions that are part of Ethereum’s design. This is because its general purpose financial infrastructure is based on: Operations are initiated and understood by humans.
Therefore, when an AI system interacts directly with a network and signs transactions without prior human review, errors are no longer conceptual, but rather leading to immediate and irreparable economic loss.
The Lambda class team’s analysis has special relevance given the circumstances of January 29th. The ERC-8004 standard has been implemented on the Ethereum main network. As reported by CriptoNoticias, this standard provides Ethereum with precisely the system that AI agents can run. Automatically connect, validate, and evaluate each other Through smart contracts.
What happens when AI replaces human operators?
According to the Lambda Class report, the library, a software toolkit that developers use to interact with Ethereum and send transactions, It is designed for humans, not for autonomous systems.
Tools like ethers.js and web3.js assume that someone understands the signature before approving a transaction. As mentioned above, this model can fail if the operator is an AI.
- agent Possibility of hallucinating the addressThis means that a valid but incorrect address will be generated.
- can confuse units. For example, “send 100 as 100 Ether instead of $100”.
- Also, manipulated Instruction injection is a technique that introduces malicious commands into the data being processed.
Each of these errors is unlikely to occur on its own. However, the report warns that these failures can occur when millions of automated trades are executed. they become unavoidable.
Ethereum does not have banks that reverse operations. Once the transaction is confirmed, funds are lost forever (Except for the famous DAO hack).
Lambda Class emphasizes that this is not a matter of “improving AI.” Tolerating imperfect systems creates risk operate directly On irreparable financial infrastructure. When something fails, the system returns technical messages that the AI cannot safely interpret.
The report describes this scenario as Robot drives truck without automatic braking: The problem is not the agent’s intentions, but the lack of barriers to stop the agent when something goes wrong.
Limitations as a layer of defense
To address this issue, the Lambda Class team believes that the way to reduce risk is not by making the AI “smarter”; Set structural limits.
To that end, he developed a development kit called eth-agent. Introduce mandatory restrictions Performing transactions in each wallet. For example, per-transaction, hourly, and daily spending limits. In this way, if the agent tries to exceed these limits, the operation will fail automaticallythere is no possibility of evasion.
The system also returns clear, structured errors. Instead of technical messages that are difficult to interpret, you’ll be told which rules were violated and when it’s safe to try again.
Additionally, for sensitive transactions (such as large amounts or new recipients) requires human approval before carrying out the shipment.
There are ways to avoid AI risks
As part of its predictions, the study advises that self-employed agents primarily operate through the following means: stable cointo avoid errors caused by price fluctuations.
We also recommend incorporating smart accounts based on the ERC-4337 standard. Delegate authority in a limited and controlled manner.
The core ideas of these proposals are similar to those of operating systems. The application may crash, but the core impose rules to prevent further damage. In decentralized finance, even if AI makes a mistake, its “core” must be protected.
The report concludes that while AI agents will continue to improve, they will never be perfect. In a financial system without error recovery, relying on its correction is insufficient.