Morning Edition · Wednesday, August 5, 2026Published at 1:32 AM EDT · New York

Researchers Report That Agent Models Internally Register When They Have Been Prompt-Injected

A preprint argues the hidden states of agentic language models carry a detectable signal of exposure to malicious instructions hidden in tool output, even when the agent goes on to obey them.

Indirect prompt injection remains the unsolved failure mode of production agents. An agent reads a web page, a document or a tool response, that content contains instructions, and the model treats them as if the user had written them. Every…