Definition
A class of vulnerability in which sensitive information from one AI agent's tasks — such as memory contents, tool outputs, or credentials — spills over to a different agent in the same system, often due to shared identifiers, insufficient isolation, or insecure file paths between agents.
Why it matters
As organisations chain multiple AI agents together to automate complex workflows, a weakness in one agent's data boundaries can expose confidential information from entirely unrelated business processes handled by other agents in the same platform.