What it does
Obvious commands are handled locally; ambiguous requests can be allowed, denied or sent for approval.
Agent tooling
Reward Hack Guard
Keeps coding agents away from graders, hidden tests and evaluation machinery while allowing ordinary development work.
Bookmark: Reward Hack Guard
Keep this in your collection.
Leave a noteWhat would you try with this?
: Reward Hack Guard
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
Ask your developer to compare two actions with this tool: printing a harmless message and changing a test's expected answer. TypeSafe, an outside AI service, checks those actions. It supplies a software access key your developer needs to connect the tool. Look at which action gets blocked, but keep reviewing allowed actions too: the AI can miss harmful ones.