Skip to content

visible UI and forms might have fewer open-ended prompt injection attacks #314

Description

@npdoty

It seems like the spec has considered a couple approaches: registering JavaScript tools, which have names and descriptions communicated just to the user's agent, and annotating/marking up forms.

While prompt injection with natural language will be a rampant problem no matter what, it seems like a safer direction might be to put instructions and intentions of tools into the visible UI, rather than the method of registering tools invisibly to the user. If the AI agent helps the user take action through the visible controls, the user can maybe see whether something is misleading, or if it's the same surface, at least having it visible makes it easier for some human user to notice the attack.

That might also have accessibility improvements, so that functionality isn't hidden from visitors who aren't using AI agents, for example.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions