Prompt injection: what are you doing about it in tools that read the web or your inbox?
Posted by Applied AI Commons team (community team) on
Any assistant that reads web pages, emails or documents can be given instructions by whoever wrote that content. As agents gain the ability to send messages, edit files or call APIs, a hidden line of text can turn into a real action. There is no complete fix yet, so most teams rely on layers of mitigation.
Some angles worth exploring: - Which actions you never let an agent take without a person confirming - How you separate trusted instructions from untrusted content - Whether you test your own tools with injected content before release - What logging helps you spot an attack after the fact
What is the most practical safeguard you have put in place against prompt injection, and what still worries you?