Insecure Inter-Agent Communication: When Agents Talk, Attackers Listen (ASI07)

Short summary

OWASP's ASI07 covers vulnerabilities in inter-agent communication where messages between AI agents lack authentication, integrity verification, or semantic validation. Real-world attacks include Trustwave's Agent-in-the-Middle exploit on Google's A2A protocol and Unit 42's agent session smuggling. The core problem mirrors early microservices mistakes: internal traffic is trusted by default, but agent messages are untyped natural language that can't be reliably validated against injection.

• •Inter-agent messages lack authentication and integrity checks, enabling interception, spoofing, and replay attacks
• •Real exploits demonstrated on Google A2A protocol and MCP endpoints via agent card poisoning and session smuggling
• •Transport encryption alone doesn't solve the problem — the natural-language message content itself is the attack vector

Generated with AI, which can make mistakes.