AI Agent Security: Why Refusals Are Not Access Controls

Hostile documents can steer a tool-using AI toward actions its user never requested. Permissions and meaningful approval checks limit the consequences.
This is the summary The Defiant publishes in its feed. The full article lives on their site.