The authors separate computer-use into action generation, which chooses a semantic action, and visual grounding, which converts that choice into clicks, keystrokes or other GUI commands.
Secure transactions limit untrusted influence on computer-use agents
The system precommits how interface content may affect each action, while grounding commands against a masked view.
Big Tech
Sarthak Choudhary · Mihai Christodorescu · Ashish Hooda · Somesh Jha · Tongxin Li · Damien Octeau
University of Wisconsin–Madison · Google · Google DeepMind
Research Digest··2 min read
Choudhary et al.
Why this paper
From Google and 2 others
In one line
Secure-CUA protects computer use agents from untrusted interface content by committing to action transactions before accessing it.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§