EchoLeak and the new security perimeter: the two OWASP lists every AI owner should know

In short: EchoLeak showed that one email with hidden instructions could make a correctly configured enterprise assistant send confidential data out, with no click. Language models treat any text they read as possible instructions, so the defense is design: least privilege, untrusted retrieved content, limited outbound channels and validated output, organized around OWASP’s two Top 10 lists.^1,2,3

Conventional software keeps code and data apart: developers write instructions, users supply data, and security lives at the boundary. A language model erases that boundary. It takes instructions from text, and text arrives from everywhere: the user, a retrieved document, an email, a web page, a tool’s output, another agent. Any of them can carry an instruction the model will follow.