Skip to main content

Decision Walk: An Email Summarizer Meets a Poisoned Thread

Apply prompt-injection controls step by step when hostile instructions appear inside email content.

ai-security-safety5 min read

The setup A forwarded email tries to convince the assistant to reveal hidden prompts before it drafts a reply. The danger is not the wording alone. It is that the same system can read content and act on it. The principle Treat messages as evidence, not authority The safe path keeps the original task intact, narrows action rights, and escalates when the content asks for secrets or role changes. Untrusted content can inform the answer, but it cannot rewrite the assistant’s operating instructions. Shortcut Argue with the malicious email in the same context and hope the model stays anchored. The…

Read the full lesson

Sign up free: one personalized lesson every day, matched to your role and goals.

← Back to library

Contact us