A team wraps fetched web content in <untrusted> tags and tells the model to ignore any instructions inside them. They consider prompt injection solved. What is the flaw in this reasoning?
Choose an answer to check it. Afterward, select any option to review its explanation.
Model-level defenses are probabilistic; least privilege and enforced controls bound the damage when an injection lands.
No agent reading untrusted content is fully immune; delimiters alone are not a security control.
There are no reserved tag names; the issue is that any delimiter is a soft boundary.
Injections commonly arrive through fetched pages, documents and tool results.
id qe0d3a81f12