77. A team proposes blocking a list of suspicious phrases to prevent prompt injection. What would you ask them to validate?
MediumApplication & data security
Technical
Show model answer
Ask which threat the filter addresses and how it behaves on legitimate and hostile content. Text matching can be a signal, but it does not establish that every instruction hidden in external content will be recognized.
Require permission enforcement, restricted tools and review at consequential action boundaries. Test whether an unrecognized instruction can cause harm despite the filter, and monitor false positives without claiming that a text scan establishes trust.
What interviewers listen for
- Define the filter's limited purpose
- Enforce permissions independently
- Test consequences beyond phrase matching