AI security
A Hidden Instruction Can Rewrite Your AI Summary
A contractor’s report could tell an assistant to hide a failed inspection. Test whether your summary preserves the finding before letting it close the job.

In a 2023 study, Kai Greshake and colleagues tested the Bing Chat sidebar in Microsoft Edge. They put instructions in parts of local webpages that the sidebar could read but that the browser did not display. In one test, they directed Bing Chat to stop using The New York Times. When asked to summarise US headlines, it left the paper out of the summary even though the links included it.
The researchers call this indirect prompt injection. An instruction reaches the assistant through a source it was asked to read. Whoever put the text there can influence the reply without taking over the user’s account. The experiments showed what could happen in the tested version, leaving the frequency of such failures in other systems open.


