AI trading agents are only as trustworthy as their data

Imagine a major crypto exchange declaring insolvency out of the blue. In the past, hackers or fraud caused wipeouts worth billions of dollars, but today? AI could just as easily be the culprit.

With AI agents that can autonomously trade on cryptocurrency exchanges being pushed by various players in the industry, agents causing a crypto crash is a plausible scenario.

Image credit: Arsal Ysfin

Simply put, if an AI agent is designed to make trades based on market information – including news articles or social media posts – it would be relatively easy to “poison” those sources with false narratives. This could trigger a wave of automated selling from agents that couldn’t distinguish the rumor from reality, which could then crash a coin or a whole market.

While no such attack has happened yet, the conditions for one already exist. The question is no longer if an AI-driven financial crisis will occur, but when – and, more unsettlingly, how little capital it might take to trigger one.

In my work as an advisor to Web3 companies and government organizations, I have watched the narrative around AI in crypto shift from cautious optimism to uncritical adoption.

Today, 45.7% of platform interactions on Binance are system-triggered rather than user-initiated, which means they are carried out by a computer, not a human. That share is only growing, and every percentage point represents a wider attack surface for anyone looking to exploit these systems.

How AI trading agents work

While AI trading agents are designed to bring efficiency, they are also highly vulnerable. The combination of autonomous agents, high-frequency trading infrastructure, and an information ecosystem saturated with synthetic media has created a perfect storm for potential attacks.

At a basic level, these agents ingest market data – price movements, order books, news, and social sentiment – and use machine learning models to identify patterns or signals that inform trading decisions. Once certain conditions are met, they execute trades automatically, often at high speed and without human intervention.

See also: The crypto label is becoming a liability for winners in the space

However, recent research underscores how fragile these agents are in ways that should alarm anyone using them.

A study released in February tested 13 AI trading models using distorted or misleading market information. Most didn’t adapt at all, and their performance barely changed, suggesting they were just following fixed strategies rather than reacting to new signals.

When false signals were introduced, some models saw sharp drops in performance, showing how easily they could be thrown off by bad data.

Plan of attack

Protection

What this means for industry players