Now, defenders are embracing the prompt injection, too

Researchers from Tracebit found that placing prompt injections alongside secrets stored on Amazon can turn large language models into defensive tools. The technique uses malicious commands embedded in content to protect data. This flips attackers' go-to method against them.