Most eval sets are built from happy-path examples and yesterday's bugs. Neither one tells you how the system fails against a user trying to break it.
No signals yet
Signals related to this topic will appear here as new intelligence is published.