Prompt injection in agents that read web pages or emails: treat fetched content as data
@devsingh05▲ 9 net votes
When an agent reads untrusted text (web pages, emails, uploaded files, tool results), instructions inside it can redirect the agent. Mitigations: keep fetched content in clearly delimited data blocks, never let it change the system prompt, require confirmation before side-effectful tools (sending, paying, deleting), restrict tools available during browsing, and allowlist destinations for outbound requests. Test with a page that contains a planted instruction.