Definition
A risk condition identified by the Cloud Security Alliance where an AI agent simultaneously has three dangerous properties: access to sensitive data, the ability to receive instructions from untrusted external sources, and the ability to take actions in the outside world (e.g. send emails, call APIs). Any agent with all three conditions is highly vulnerable to being manipulated into causing serious harm.
Why it matters
Research found that 98% of production AI agents already carry all three conditions, yet only 11% have adequate defences. This means the vast majority of deployed enterprise agents are structurally one crafted malicious input away from a significant incident.