Mahal Systems

Overview › Benchmark catalog

macOSWorld

Fits agents that must run on Mac desktops or in several languages. Borrow the multilingual instruction and interface variants and the deception subset.

What it asks the agent to do

  • complete tasks in 30 macOS apps (system, productivity, media)
  • follow instructions in English, Chinese, Arabic, Japanese or Russian
  • resist on-screen deception attacks in the safety subset

What to borrow from it

sandbox designtaxonomy

Caveats

Needs an AWS macOS instance and 15 to 20 minutes per task; grading details not confirmed from abstract or README; license not stated.

Details

Workloadscomputer use, safety
Interactionsingle turn
Environmentdesktop
End state checkedYes
Toolscloud hosted macOS instance, mouse and keyboard
Gradingprogrammatic check
Size202
Datasynthetic
Personal datanone
Licenseunknown
Harnessown repo
Statusactive

Links