AI Crypto Plugins Pose Security Risks as Vulnerabilities Exposed

Published: Sun 25 May 2025 | 5 min read | By Mikaia A.

Key Takeaways

• Crypto AI agents use the Model Context Protocol (MCP), which is flexible but vulnerable to targeted attacks.
• Malicious plugins can hijack AI agents to steal private keys and crypto funds.
• SlowMist identified four major attack vectors through an educational project called MasterMCP.
• Securing plugins, behaviors, and privileges must become a top priority for crypto AI developers.

The Emergence of a New Threat

Artificial intelligence is rapidly entering the crypto space. By the end of 2024, over 10,000 crypto AI agents were active, with projections exceeding one million by 2025. These AI agents, seen as a revolution in the sector, are not standalone models like GPT-4 but extensions connected in real time to wallets, bots, or dApps.

Their mission? To make automated decisions and execute on-chain actions. All based on the Model Context Protocol (MCP). However, this flexibility is also its weakness. MCP acts as the brain of these agents, deciding which tools to use and how to respond. According to SlowMist, this architecture opens an "uncontrollable surface without strict sandboxing." Malicious plugins can hijack an agent, inject toxic data, or make it call trapped external functions.

Security expert Monster Z explains:

"Poisoning of agents and MCPs results from malicious information introduced during the interaction phase."

In short, even a well-trained agent can betray if it receives a toxic instruction at the wrong time. Worse: this threat surpasses classic AI model poisoning in severity.

A System That Can Self-Destruct from Within

The attacks are diverse, precise, and sneaky. SlowMist documents four main ones in its report. The MasterMCP project reproduces them to help developers understand the danger.

1. Data poisoning uses plugins to make the agent perform absurd tasks or mislead the user.