Cyber Risks · Unit 42 Palo Alto Networks

When AI Agents Go Rogue: Agent Session Smuggling Attack in A2A Systems

Summary: Unit 42 Palo Alto Networks disclosed a new attack technique called agent session smuggling, in which a malicious AI agent exploits an established cross-agent communication session under the Agent2Agent (A2A) protocol to inject covert instructions between a legitimate client request and the server's response. In two proof-of-concept scenarios, a malicious research assistant agent induced a financial assistant agent to disclose its system configuration, tool schemas and entire session history, and to execute an unauthorized trade of 10 shares without the user's knowledge.

Bottom line: The researchers note they have not observed the attack in the wild, but say the barrier to execution is low: an adversary needs only to convince a victim agent to connect to a malicious peer, after which the injected instructions would remain invisible in a typical production chatbot interface that shows only the final response.

Read the original at Unit 42 Palo Alto Networks ↗

Topics
Malware
Related