New Protocol Vulnerabilities Allow Malicious Prompts to Spread Between AI Agents

Security vulnerabilities in a new communication protocol are allowing attackers to pass malicious instructions between artificial intelligence agents inside corporate networks, Ars Technica reported on October 5, 2026. The technique targets individual agents to exfiltrate database contents alongside sensitive business and personal information. [1]

According to Ars Technica, Google and four other organizations have acknowledged vulnerabilities over the past five months that exploit these trust gaps. The attacks use a specialized form of prompt injection that targets specific internal agents, including those used for data analysis or translation tasks. [1]

The security risks stem from the rapid adoption of AI agents across millions of organizations worldwide, Ars Technica reported. Attackers leverage these interconnected systems to spread harmful commands across internal networks once an initial agent is compromised. [1]