AI Agents: A New Frontier for Cyberattacks

The widespread adoption of AI agents in organizations is introducing novel security risks, creating avenues for attackers to compromise systems. A newly identified attack class, termed 'protocol pivoting,' exploits trust vulnerabilities between interconnected AI agents.

This method involves an attacker compromising one agent and using it to send malicious instructions to others within a network. Independent researcher Syed Anas Mohiuddin demonstrated how these exploits leverage gaps in the Model Context Protocol (MCP), a standard for inter-agent communication. Attacks can lead to sensitive data exfiltration or server-side request forgery.

Many specialized AI agents lack robust security guardrails, and their inherent trust in other internal agents allows exploits that would otherwise be blocked. The vulnerability in Google's systems, rated 8 out of 10 for severity, stemmed from an inadequately configured HTTP client. Google has since implemented IP range allow-lists and block lists as a fix.

Researchers note that these attacks are a form of indirect prompt injection, exploiting trust assumptions across different communication protocols. The issue arises because organizations are rapidly deploying AI architectures without adhering to the 'zero trust' security model, which assumes no node can be implicitly trusted. A critical takeaway is to treat all information passed from AI agents as potentially hostile input.