A pattern: a proof-of-human for un-quarantining feedback text.
- A pattern: a proof-of-human for un-quarantining feedback text.
- My little Floored Fitness app has a feature to allow users to add feedback.
- I use it to give directives to my agents to add features, "programming" mostly by adding feedback.
- Feedback from me (the admin) is safe, but feedback from anonymous third parties could be dangerous.
- It could, for example, contain prompt injection.
- That's why non-admin feedback goes into a quarantine, that a human has to drain and approve, after verifying there's no prompt injection in the input.
- My agents were so eager to get at the 3rd party feedback that they looked at it themselves to verify it didn't have prompt injection, potentially tainting themselves.
- D'oh!
- I added a feature to the feedback tool: the human has to authenticate with TouchID before being able to see the quarantined feedback.
- That helps prove that you're a human before seeing the potentially prompt injected input.